You are on page 1of 52

L1 Adaptive Control for Safety-Critical Systems

GUARANTEED ROBUSTNESS WITH FAST ADAPTATION

NAIRA HOVAKIMYAN, CHENGYU CAO, EVGENY KHARISOV, ENRIC XARGAY, and IRENE M. GREGORY

S

Digital Object Identifier 10.1109/MCS.2011.941961 Date of publication: 16 September 2011

afety-critical systems appear in several application areas, such as transportation and air-traffic control systems, nuclear plants, space systems, and operating rooms in hospitals. Reliable control of these systems requires not only meeting performance specifications in the presence of multiple constraints, which together ensure predictable response of the overall system and safe operation, but also graceful performance degradation when the
1066-033X/11/$26.00©2011IEEE

54 IEEE CONTROL SYSTEMS MAGAZINE

»

OCTOBER 2011

NASA LANGLEY/SEAN SMITH

underlying assumptions are violated. Figure 1 explains this requirement for flight control applications. The green region for angle of attack and sideslip represents the normal flight envelope, where the airplane usually flies in the absence of abnormalities. The light blue area represents configurations for which high-fidelity nonlinear aerodynamic models of the aircraft are available from wind-tunnel data. Outside this wind-tunnel data envelope, the aerodynamic models available are typically obtained by extrapolating wind-tunnel test data and hence are highly uncertain. This fact suggests that pilots might not be adequately trained to fly the aircraft in these regimes. Moreover, it is not reasonable to rely on a flight control system to compensate for the uncertainty in these flight conditions, since aircraft controllability is not guaranteed in such regimes. The main objective of the flight control system therefore, from safety considerations, is to ensure that an aircraft, suddenly experiencing an adverse flight regime or an unexpected failure, does not “escape” its a– b wind-tunnel data envelope, provided that enough control authority remains. This objective requires that the control system quickly adapt to the failure with guaranteed and uniform transient performance specifications to ensure the safety of the aircraft. Typical performance specifications in control applications include transient and steady-state performance, as well as robustness margins that the control engineer must be able to trade off in a systematic way subject to hardware constraints, such as CPU, sampling rates of sensors and actuators, and control channel bandwidth. This viewpoint has led to certification of flight control laws for commercial aviation, where the certification protocols rely on the gain and phase margins of the gain-scheduled controllers computed for all operating points [1]. This process is repeated for each aircraft, rendering the overall verification and validation (V&V) expensive. The price of this process increases with growing system complexity. L1 adaptive-control theory is motivated by the emerging need to certify advanced adaptive flight critical systems with a more affordable V&V process. On the one hand, this objective requires the development of a control architecture with a priori quantifiable transient and steady-state performance specifications and robustness margins. On the other hand, achieving this objective appears to be possible with an architecture that enables fast and robust adaptation with uniform performance bounds [2, Def. 4.6] without losing robustness. In this context, fast adaptation indicates that the adaptation rate in L1 architectures is to be selected so that the time scale of the adaptation process is faster than the time scales associated with plant parameter variations and the underlying closed-loop dynamics. Robust adaptation indicates that, despite fast adaptation in L1 architectures, the robustness properties of the closed-loop adaptive system can be adjusted independently of the adaptation rate. Because the emphasis is on uniform performance bounds for transient

50 40 α (deg) 30 20 10 0 –40 –30 –20 –10 0 10 β (deg)

20

30

40

Normal Flight Envelope Current Wind Tunnel Data Loss-of-Control Accident Data FIGURE 1 Loss of control accident data relative to angle of attack a and sideslip b (from [77]). Angle of attack a and angle of sideslip b are two of the state variables describing aircraft dynamics. The green region is the combination of these variables associated with the normal flight envelope. The light blue area corresponds to configurations for which there are accurate models based on wind-tunnel data. The gray area corresponds to configurations for which the aerodynamic data are extrapolated from the light blue area and are highly uncertain. This fact suggests that pilots might not be correctly trained to fly the aircraft in these a2b conditions and can potentially cause dangerous oscillations, known as pilot-induced oscillations.

and steady-state operation, a sufficient condition for stability and performance is derived in terms of L1 -norms of the underlying transfer functions, which leads to uniform bounds on the L` -norms of the input-output signals. Therefore, the underlying theory is named L1 adaptive-control theory. For the definition and properties of the L1 -norm of a system see “ L1 -Norm of a System.” The key feature of L1 adaptive-control architectures is guaranteed robustness in the presence of fast adaptation, which leads to uniform performance bounds both in transient and steady-state operation, thereby eliminating the need for gain scheduling of the adaptation rates [3]. These properties can be achieved by appropriate formulation of the control objective with the understanding that uncertainty in a feedback loop cannot be compensated outside of the control channel bandwidth. By explicitly building the robustness specification into the problem formulation, it is possible to decouple adaptation from robustness and increase the speed of adaptation, subject only to hardware limitations. With L1 adaptive-control architectures, large learning gains appear to be beneficial both for performance and robustness, while the tradeoff between the two is resolved by selecting the underlying filter structure. The latter is a linear problem and thus can be addressed using conventional methods from classical and robust control. Moreover, the performance bounds of L1 adaptive-control
OCTOBER 2011

«

IEEE CONTROL SYSTEMS MAGAZINE 55

L1-Norm of a System
he L1-norm of a system sets the relation between the peak values of the system’s input and output. The L1-norm is also called the peak-to-peak gain of a system. Let G 1 s 2 be a proper and exponentially stable system. Assume zero initial conditions. Then, for the bounded input u 1t 2 , its output y 1t 2 can be written as

T

`

7 G 1s 2 7 L1 ! 3 0 g 1 t2 0 d t,
0

which leads to the bound 7 y 7 L` # 7 G 1 s 2 7 L1 7 u 7 L` . (S1)

t

y 1t 2 5 g 1t 2 * u 1t 2 5 3 g 1 t 2 t 2 u 1 t2 d t,
0

where “*” denotes the convolution operator, and g 1t 2 is the impulse response of G 1s 2 . Letting 7 y 7 L` ! sup 0 y 1t 2 0 ,
t$0

Notice that the bound in (S1) holds if and only if the system G 1 s 2 is exponentially stable and proper [4]. For unstable or improper systems the L1-norm does not exist, since the impulse response is unbounded. For the m -input, l -output exponentially stable, proper system G 1 s 2 , the L1-norm is defined by 7 G 1s 2 7 L1 ! max a a 3 0 gij 1 t 2 0 d tb , i 5 1, c, l
j51 0 m `

we obtain the bound
t

0 y 1t 2 0 5 ` 3 g 1 t2 u 1t 2 t 2 dt `
0 t

# 3 0 g 1 t2 0 0 u 1t 2 t2 0 d t
0 `

where gij 1 t 2 is the 1 i, j 2 entry of the impulse-response matrix of the system G 1 s 2 . If G1 1 s 2 and G2 1 s 2 are exponentially stable proper systems, then 7 G 1 s 2 7 L1 # 7 G2 1 s 2 7 L1 7 G1 1 s 2 7 L1, where G 1 s 2 5 G2 1 s 2 G1 1 s 2 [4].

# 3 0 g 1 t20 d t 7 u 7 L`.
0

The L1-norm of G 1 s 2 is defined as

architectures can be analyzed to determine the extent of the modeling of the system required for the given hardware. This article uses a scalar example to explain the key concept of this theory in terms of decoupling adaptation from robustness. Additional details are given in [4]. We also describe extensions for a broader class of systems involving unknown input gain and unmodeled dynamics. A simpler version of this analysis is presented in [5] for a scalar example. Here we extend the discussion from [5] to systems of arbitrary dimension with unknown input gain. Two benchmark examples are used to illustrate robustness and performance tradeoffs, Rohrs’s example, and the two-cart system. For more details we refer the reader to [4] and [6]. Key theorems and lemmas that support the theoretical results are stated in sidebars with appropriate references, without the details of proofs. Several flight test examples are included as real-world applications. Additional detail on the flight test results is given in [7] and [8].

FROM THE “BRAVE ERA” TO L1 ADAPTIVE CONTROL
Research in adaptive control was motivated in the 1950s by the design of autopilots for highly agile aircraft that need to operate over a wide range of speeds and altitudes, experiencing large parametric variations. In the early 1950s, adaptive control was conceived and put forward as a technology for automatically adjusting the controller parameters in the face of changing aircraft dynamics [9], [10]. In [11], that
56 IEEE CONTROL SYSTEMS MAGAZINE

period is called the brave era because “there was a very short path from idea to flight test with very little analysis in between.” The tragic flight test of the X-15 confirms this view [12]. The initial results in adaptive control were motivated by system identification [13], which led to an architecture consisting of an online parameter estimator combined with automatic control design [14], [15]. Two architectures of adaptive control emerged, namely, the direct method, where controller parameters are estimated, and the indirect method, where process parameters are estimated and the controller parameters are obtained using a design procedure. The relationships between these architectures are clarified in [16]. Progress in systems theory has led to fundamental results for the development of adaptive-control architectures [16]–[26]. Along the same lines, Rohrs’s example challenged the robustness of adaptive controllers in the presence of unmodeled dynamics [27]. Although [27] includes a proof of the existence of two infinite-gain operators in the closed-loop adaptive system, the explanation given for the phenomenon observed in the simulations, which was based on qualitative considerations, was not complete. Further details can be found in [28] and [29]. Nevertheless, [27] emphasizes a key point, namely, the available adaptive-control algorithms to that date were not able to limit the bandwidth of the closed-loop system and guarantee its robustness. The results and conclusions of [27] motivated numerous investigations of robustness and

»

OCTOBER 2011

stability issues of adaptive-control systems. In [30]–[35], the causes of instability are analyzed, and damping-type modifications of adaptation laws are suggested to prevent them. The basic idea of these modifications is to limit the gain of the adaptation loop and eliminate its integral action. Examples of these modifications are the s -modification [32] and the e -modification [35]. Although these modifications address the problem of parameter drift, they do not directly address the architectural problem identified in [27]. An overview of robustness and stability issues of adaptive controllers can be found in [29]. In adaptive control, the nature of the adaptation process plays a central role in both robustness and performance. Ideally, adaptation is expected to correctly respond to all changes in initial conditions, reference inputs, and uncertainties by quickly identifying a set of control parameters that provide a satisfactory system response. This fact demands fast estimation schemes with high adaptation rates and, as a consequence, leads to the fundamental question of determining an upper bound on the adaptation rate that does not result in poor robustness characteristics. We notice that the results of [36, p. 549] limit the rate of variation of uncertainty by providing examples of destabilization due to fast adaptation, while the transient performance analysis is reduced to a persistency of excitation assumption [37], which cannot be verified a priori. The lack of analytical quantification of the relationship between the rate of adaptation, the transient response, and the robustness margins led to gain-scheduled designs of adaptive controllers, examples of which are the flight tests of the late 1990s by the U.S. Air Force and Boeing [38], [39]. These flight tests relied on intensive Monte Carlo analysis for determining the “best” rate of adaptation for various flight conditions. It was apparent that fast adaptation led to high frequencies in control signals and increased sensitivity to time delays. The fundamental question was thus reduced to determining an architecture that would allow for fast adaptation without losing robustness. It was understood that this architecture can reduce the amount of gain scheduling and possibly eliminate gain scheduling, since fast adaptation—in the presence of guaranteed robustness—can compensate for the negative effects of rapidly time-varying uncertainty on the system response. L1 adaptive-control theory addresses this question by setting in place an architecture for which the estimation loop is decoupled from the control loop. This decoupling allows for an arbitrary increase of the estimation rate, limited by only the available hardware, that is, the CPU clock speed, while robustness is limited by the available control channel bandwidth and can be addressed by conventional methods from classical and robust control. The architectures of L1 adaptive-control theory have guaranteed transient performance and guaranteed robustness in the presence of fast adaptation, without introducing or requiring persistence of excitation, without gain scheduling of

the controller parameters, without control reconfiguration, and without resorting to high-gain feedback. With an L1 adaptive controller in the feedback loop, the response of the closed-loop system can be predicted a priori, thus reducing the amount of Monte Carlo analysis required to verify and validate these systems. These features of L1 adaptive-control theory are exhibited by flight tests and in mid-to-high fidelity simulation environments [7], [8], [40]–[60]. In the remaining sections of this article we present the two basic architectures of adaptive control, direct and indirect, and use the indirect architecture for transition to L1 adaptive control. We discuss various insights and properties by analyzing two benchmark problems, specifically, Rohrs’s example and the two-cart system. Flight tests of NASA’s Airborne Subscale Transport Aircraft Research (AirSTAR) testbed conclude the article.

LIMITATIONS AND OPPORTUNITIES INDUCED BY ARCHITECTURES
In this section we place the focus on the architecture. We first present the direct model reference adaptive control (MRAC) architecture using a scalar system. We proceed by considering its state-predictor-based reparameterization, and we preview the L1 adaptive controller. We emphasize the role of the predictor in the architecture. Consider the first-order plant # x 1 t 2 5 2amx 1 t 2 1 b 1u 1 t 2 1 u x 1 t 22 , x 1 0 2 5 x0, (1)

where x 1 t 2 [ R is the state of the system, u 1 t 2 [ R is the control input, am [ 1 0, ` 2 defines the desired pole location, b [ 1 0, ` 2 is the known system input gain, and u [ R is a constant uncertainty with the known bound 0 u 0 # u max. (2)

The control objective is to define the feedback signal u 1 t 2 such that x 1 t 2 tracks a given bounded piecewise continuous input r 1 t 2 [ R with desired performance specifications. We assume that 7 r 7 L` # r. The MRAC architecture proceeds by considering the ideal controller uid 1 t 2 5 2u x 1 t 2 1 kgr 1 t 2 , where kg ! am b (4) (3)

is the inverse of the dc gain of the plant (1), which yields unit dc gain of the closed-loop system (1), (3) with u 1 t 2 5 uid 1 t 2 . Thus, the choice of kg in (4) ensures that x 1 t 2 tracks step reference inputs r 1 t 2 with zero steady-state error. In fact, (3) provides perfect cancellation of the uncertainty in (1) and leads to the ideal system
OCTOBER 2011

«

IEEE CONTROL SYSTEMS MAGAZINE 57

To show that the tracking error converges to zero. # It follows from (7) that e 1 t 2 is uniformly bounded [2. and thus the signals e 1 t 2 . Def. To analyze the asymptotic properties of this adaptive scheme. 2 2G (9) r ∧ −θx + kgr Control Law C(s) x ⋅ x = −amx + b (u + θx) Plant ∧ ⋅ ∧ ∧ ∧ x x = −amx + b (u + θx) – ∼ x # The time-derivative V 1 t 2 of V 1 e 1 t 2 . e 1 0 2 5 0. 58 IEEE CONTROL SYSTEMS MAGAZINE which implies that e 1 t 2 S 0 as t S ` . the tracking error dynamics can be written as – e x ⋅ x = −amx + b (u + θx) Plant ⋅ ∧ θ = −Γxe Adaptation Law (a) # u 1 t 2 x 1 t 2 . u 1 t 2 5 2u (6) ^ 1 t 2 [ R is an estimate of u . the equilibrium of (7)–(8) is Lyapunov stable. we compute the second derivative $ # V 1 t 2 5 22am e 1 t 2 e 1 t 2 . Substituting (6) into (1) where u yields the closed-loop system # x 1 t 2 5 2amx 1 t 2 2 b| u 1 t 2 x 1 t 2 1 bkgr 1 t 2 . e 1 t 2 5 2ame 1 t 2 1 b| The update law for the parametric estimate is given by # ^ 102 5 u ^ 0. xm 1 0 2 5 x0. the ideal controller (3) is not implementable since this controller explicitly uses the uncertain plant parameter u in its definition. consider the Lyapunov-function candidate 1 b 2 V1e1t2. | u 1 t 2 are bounded.⋅ xm = −amxm + bkgr Ideal System r ∧ u = −θx + kgr Control Law ∧ θ u xm ^ 1 t 2 2 u denotes the parametric estimation where | u 1t2 ! u error. ^ 1 t 2 5 2Gx 1 t 2 e 1 t 2 . (5) with state xm 1 t 2 [ R. The model reference adaptive-control (MRAC) architecture with (b) state predictor is equivalent to the (a) conventional MRAC architecture. Defining the tracking error signal e 1 t 2 ! xm 1 t 2 2 x 1 t 2 . | u 1 t 22 along the system trajectories (7)–(8) is given by # ' V 1 e. # V 1 t 2 5 2ame2 1 t 2 # 0. The (c) L1 controller is based on the MRAC architecture with state predictor but has a lowpass filter C 1 s 2 in the control channel. we obtain ∧ θ State Predictor ⋅ ∧ ∼ ∼ θ = −Γxx Adaptation Law (c) FIGURE 2 Adaptive-control architectures for the scalar case. tS` # xm 1 t 2 5 2amxm 1 t 2 1 amr 1 t 2 . 5 2ame 2 1 t 2 1 b| u 1t2x1t2e1t2 1 | u 1t2u G Using the adaptation law (8). $ # 4. Hence.6]. it follows that x 1 t 2 is bounded. and the initial con^ 0 are selected accordditions for the parametric estimate u ing to (2). Since x 1 t 2 5 xm 1 t 2 2e 1 t 2 and xm 1 t 2 is the state of the exponentially stable ideal system (5). | u2 | ' V 1 e. Application of Barbalat’s lemma. u u (7) (8) r ∧ u = −θx + kgr u ⋅ = −a x + b (u + θx) x m Plant ⋅ ∧ ∧ ∧ x = −amx + b (u + θx) x Control Law ∧ x ∼ x – ∧ θ State Predictor ⋅ ∧ ∼ θ = −Γxx Adaptation Law (b) u where G [ 1 0. The architecture of the closed-loop system is given in Figure 2(a). Thus. stated in “Barbalat’s Lemma. ` 2 is the adaptation gain. and x 1 t 2 follows xm 1 t 2 as t S ` with desired specifications given with the help of the ideal system (5). | u2 # # e 1t2 1 u 1t2 V1t2 5 'e '| u # b 5 e 1 t 2 12ame 1 t 2 1 b| u 1 t 2 x 1 t 22 1 | u 1t2| u 1t2 G # b ^ 1t2. Model Reference Adaptive Control The model reference adaptive controller is obtained by replacing the unknown parameter u in the ideal controller ^ 1 t 2 yielding the implementable control (3) by its estimate u law ^ 1 t 2 x 1 t 2 1 k gr 1 t 2 . However.” yields # lim V 1 t 2 5 0. implying that V 1 t 2 is uniformly continuous. | u 1 t 22 5 e2 1 t 2 1 | u 1t2. x 1 0 2 5 x0. » OCTOBER 2011 . x 1 t 2 2xm 1 t 2 converges to zero. and hence V 1 t 2 is bounded.

0| x 1t2 0 # Taking into account that ^ 1 0 2 2u 0 # 2u max. V 1 t 2 5 2am| (13) MRAC with State Predictor Next. (15) where G [ 1 0. the convergence of the integral t f (t ) 1 i=0 i=1 i=2 i=3 i=4 t 0 FIGURE S1 Barbalat’s lemma. Then tS` 3 f 1 t 2 dt 0 to a finite number as t S ` does not imply that the function f 1t 2 S 0 as t S ` . Notice that ^ 1 t 2 2 x 1 t 2 and | where | x 1t2 ! x u 1t2 ! u (11) is identical to the error dynamics (7) and is independent of the control signal u 1 t 2 . The adaptaeter estimate u 2u max "G . Notice that convergence of the parametric estimation error | u 1 t 2 to zero is not guaranteed. LEMMA S1 [2. we can initialize the ^ 1 0 2 5 x0. as 0| x 1t2 0 # OCTOBER 2011 ^ 1 t 2 2 u . The choice of the Lyapunov-function candidate 1 2 1 2 V1| x 1t2. the function f 1t 2 is continuous but not uniformly continuous because the slope of f 1t 2 tends to infinity as i S `. Since V 1 t 2 # 0 for all t $ 0. rather than only continuity. The parametric estimation errors are guaranteed only to be bounded. | u 1 t 22 5 | x 1t2 1 | u 1t2 2 2G leads to # x 2 1 t 2 # 0. which is denoted as # V1t2 ! V1| x 1t2. (8) using the state predictor # ^ 1 t 2 x 1 t 22 . This func` tion is continuous and e0 f 1 t 2 d t 5 e. tion law (12) is similar to (8) in its structure. System (10) where x replicates the plant structure (1). c is the number of the spike. Notice that. | x 1 0 2 5 0. for all t $ 0. (11) implying that the errors | x 1 t 2 and | u 1 t 2 are uniformly bounded. is critical for convergence of f 1 t 2 . Consider the Lyapunov function (13) evaluated along the system trajectories (11)–(12). state predictor with x we obtain the prediction error dynamics # | x 1 t 2 5 2am| x 1 t 2 1 b| u 1t2x1t2. 0| u 1 02 0 5 0 u bound (15) can be rewritten. we obtain | u2102 1 |2 x 1t2 # V1t2 # V102 5 . with the unknown ^ 1 t 2 . which contains triangular spikes of equal height one and area 1/i !. from Figure S1 we see that f 1t 2 does not have a limit as t S ` . ` 2 . and the initial conditions for the param^ 0 are selected according to (2). Then f 1 t 2 is continuous and 1 3 f 1 t2 d t 5 a 5 e.Barbalat’s Lemma F or a continuous function f : R S R. (16) « IEEE CONTROL SYSTEMS MAGAZINE 59 . since parameter u replaced by its estimate u the state of the plant (1) is measured. x ^ 1 t 2 5 2am x ^ 1t2 1 b1u1t2 1 u ^ 1 0 2 5 x0 . except that the tracking error e 1 t 2 is replaced by the prediction error | x 1 t 2 .1. For example. for all t $ 0. However. x (10) ^ 1 t 2 [ R is the state of the predictor. The function f 1t 2 contains triangular spikes of equal height one and area 1/ i !. LEMMA 8. i 0 ! 0 However. In the example in Figure S1. u (12) | 0 u 102 0 "G . exists. | u 1 t 22 . u x 1t2. ` 2 and assume that t tS` 0 ` ` lim f 1 t 2 5 0 . Barbalat’s lemma invokes an additional assumption on f 1t 2 . lim 3 f 1 t 2 d t The assumption on uniform continuity. consider the function f 1t 2 shown in Figure S1. (14) 2 2G which leads. By subtracting (1) from (10). ^ 1 t 2 be given by Let the adaptation law for u # ^ 1 t 2 5 2Gx 1 t 2 | ^ 1 0 2 5 u 0. 2. f 1t 2 does not have a limit as t S ` . which ensures that f 1t 2 converges to zero as t S ` .2] Let f : R S R be a uniformly continuous function on 3 0. we consider a reparameterization of the architecture (6). where i 5 0.

However. However. applications requiring identification schemes with time scales comparable with those of the closed-loop dynamics tend to be challenging due to undesirable interactions between the two processes [29]. Therefore. aims for partial compensation of the uncertainty u x 1 s 2 . (12). the control signal serves as the input only to plant (1). including the transient phase. One of the key characteristics of a control system is the available bandwidth. u1s2 5 C1s2h (18) L1 ADAPTIVE CONTROL where C 1 s 2 is an exponentially stable. The available bandwidth of a control system refers to this frequency range. ^ 1 t 2 ! 2u h (19) The architecture of the L1 adaptive controller is shown in Figure 2(c). If we use the control law (6) in (10). the closed-loop system (1). we obtain # ^ 1 t 2 5 2amx ^ 1 t 2 1 bkg r 1 t 2 . by increasing the adaptation gain G. Thus. Figure 2(a) and (b) illustrates the fundamental difference between the MRAC and the predictor-based MRAC. v b 4 . by compensating for only low-frequency content of u x 1 s 2 within the bandwidth of the control channel. Barbalat’s lemma can be invoked to conclude that | x 1 t 2 S 0 as t S ` . (20) Tuning Challenges From (16) it follows that the tracking error can be arbitrarily reduced for all t $ 0. consider the case where the parameter u is known.Available Bandwidth of Control Systems n classical control the bandwidth of a system is defined as the frequency range 3 0. Most importantly. Moreover. x ^ 1 0 2 5 x0 . where v b is the frequency at which the magnitude of the frequency response is 3 dB less than the magnitude at zero frequency. Rather. This feature is used in [61] to obtain the L1 adaptive-control architecture. (10) with the L1 adaptive controller (18) does not behave similarly to the ideal system (5) due to the limited bandwidth of the control channel enforced by C 1 s 2 . These limitations on the ability to obtain a model of the plant impose constraints on the frequency range in which a controller can achieve performance improvement. which shows that the closed-loop state predictor replicates the bounded ideal system of (5). from the control law (6) and the adaptation laws in (8) and (12) it follows that large adaptive gains result in high-gain feedback control. As a consequence. however. tuning such applications is done by either computationally expensive Monte Carlo simulations or trial-and-error methods following empirical guidelines or engineering intuition. the available bandwidth is an a priori constraint imposed by the available plant model used for the control design. especially for high frequencies. and the presence of unmodeled I dynamics in the plant limit the fidelity of the frequency response model of the plant. that without introducing the feedback signal u 1 t 2 we cannot apply Barbalat’s lemma to conclude ^ 1 t 2 can diverge convergence of | x 1 t 2 to zero. Notice. the controller in (18) takes the form of the reference controller uref 1 s 2 5 C 1 s 2 1 kgr 1 s 2 2 u xref 1 s 22 . the available bandwidth does not depend on the compensator. x (17) adaptation gain. strictly proper low^ 1 s 2 is the Laplace transform of the signal pass filter. as compared to the ideal control law (3). Both x 1 t 2 and x at the same rate. efficient tuning of MRAC architectures represents a major challenge. control signal (6) can be redefined without affecting the proof of stability of the prediction error dynamics (11). namely. in the predictor-based MRAC (10). the plant and the predictor. Then. The frequency v b is also called the cutoff frequency. keeping | x 1 t 2 uniformly bounded. The digital nature of control implementations. Due to the lack of systematic design guidelines for selecting an adequate 60 IEEE CONTROL SYSTEMS MAGAZINE Notice that this control law. which manifests itself by high-frequency oscillations in the control signal and reduced tolerance to time delays. The architecture of the closed-loop system with the predictor is given in Figure 2(b). the control signal is provided as the input to both systems. while h ^ 1 t 2 x 1 t 2 1 k gr 1 t 2 . Comparing the closed-loop state predictor (17) with the ideal system (5) and the error dynamics (7) with (11). while in (a). the available bandwidth is finite [14]. which is defined as the frequency range over which the unstructured multiplicative perturbations are less than unity [14]. To derive the dynamics of the reference system for the L1 controller. Hence. in (b). The L1 adaptive controller is obtained from the predictorbased MRAC by letting the control be given by ^ 1s2. we see that the state-predictor parameterization of MRAC is equivalent to the MRAC architecture. Substituting the reference controller (20) into the plant dynamics (1) leads to the L1 reference system xref 1 s 2 5 H 1 s 2 C 1 s 2 kgr 1 s 2 1 H 1 s 2 1 12C 1 s 22 u xref 1 s 2 1 xin 1 s 2 . sensor and actuator uncertainties. Unlike the predictor-based MRAC. (21) » OCTOBER 2011 .

which corresponds to the desired behavior of the system in the absence of uncertainty. The L1 reference system can be equivalently represented in this form. objective. compensation of only the low-frequency content of the uncertainty u x 1 s 2 within the bandwidth of the control channel. 40. we obtain C 1s 2 u x 1s 2 5 C 1s 2 x 1s 2 2 C 1s 2 u 1s 2 . In the frequency domain. Portland. 1. which yields xin 1s 2 . This structural similarity of these architectures facilitates application of robust control design methods to the design of the filter C 1 s 2 of the L1 Substituting (S3) into the equation for the reference control signal (20) and isolating u 1s 2 . vol. The first term in (21) contains the ideal system (5) and the filter. 1993. pp. we can rewrite (1) as x 1s 2 5 H 1s 2 1u 1s 2 1 u x 1 s 22 1 x in 1 s 2 . which assumes that the control system encapsulates either implicitly or explicitly a model of the system to be controlled. 116. 0. Jpn. namely. and H1s2 ! b . Then assuming that 7 G 1 s 2 u 7 L1 . (25) L1 OCTOBER 2011 IEEE CONTROL SYSTEMS MAGAZINE 61 . Electron. (12). Notice that all elements of the controller implementation in Figure S2 are proper and stable. where xref 1 t 2 [ Rn is the state. 5. OR. 1987. Tomizuka. C. 473–485. Hovakimyan. s 1 am xin 1 s 2 ! 1 x . Aug. multiplying (S2) by C 1s2 and then dividing by H 1s 2 . More details about the bandwidth of the control system are given in “Available Bandwidth of Control Systems. Soc. the ideal system and the plant within the frequency range specified by the bandwidth of the lowpass filter C 1 s 2 . similar to disturbance observers. vol. no. (8)..” in Proc. [S3] W.. Disturbance observer control design is based on the internal model principle. 2011. A block diagram of the closed-loop L1 reference Notice that 1 1/12C 1 s 2 2 1 C 1 s 2 /H 1 s 2 2 is proper for all strictly adaptive controller [S4]. vol. REFERENCES system is shown in Figure S2. Contr.” ASME J. 1994. [S2] T. With this assumption. Kim.. Elect.Bridging Adaptive and Robust Control C onsider plant (1). H 1s 2 (S3) C (s) Reference H(s) Controller FIGURE S2 Reference controller for the L1 architecture. analyzed in [S1]–[S3]. s 1 am 0 We notice that xin 1 s 2 is the Laplace transform of the ideal system’s response to the initial condition.-K. we consider zero initial conditions. [S4] E. we obtain the bound 7 xref 7 L` # 7 H 1 s 2 C 1 s 2 kg 7 L1 7 r 7 L` 1 7 H 1 s 2 1 12C 1 s 22 u 7 L1 7 xref 7 L` 1 7 xin 7 L`. K. pp. 83–86. 39–44. Ohnishi. Hori. This representation is structurally similar to the disturbance observer architecture considered in [S1]–[S3]. the bound (22) can be solved for 7 xref 7 L` to obtain 7 xref 7 L` # 7 H 1 s 2 C 1 s 2 kg 7 L1 7 r 7 L 1 7 xin 7 L ` ` (22) (23) (24) 12 7 G 1 s 2 u 7 « . The L1 reference controller. Let G 1 s 2 ! H 1 s 2 1 12C 1 s 22 . we obtain the reference controller C 1s 2 C 1s 2 1 kg r 1s 2 2 xref 1s 2 . Yang and M.” IEEE Trans. (10). The L1 adaptive controller pursues a less ambitious.” Notice. Eng. AIAA Guidance. pp. “Limiting behavior of L1 adaptive controllers. This representation of the L1 reference system is structurally similar to the disturbance observer architecture. “A new servo method in mechatronics. 107-D. and N. Kharisov. This approach differs from the MRAC schemes (6). yet practically achievable. “Robust servo system design with two degrees of freedom and its application to novel motion control of robust manipulators. AIAA-2011-6441.. X. Taking the L1-norm of the transfer functions in (21).. Dynamic Syst. (S2) r uref C (s) – – xref ⋅ xref = −amxref + b (uref + θxref) C (s) Plant For simplicity of explanation. Meas. Umeno and Y. Wang. Ind. that the consequence of the lowpass filter in the control channel is that the stability of the L1 reference system is not guaranteed a priori as it is for the ideal system (5). Navigation and Control Conf. The second term depends on the uncertainty u x 1 s 2 . where the ideal controller (3) attempts to follow the ideal system and compensate for the uncertainty u x 1 s 2 in the entire frequency range. compensates for the mismatch between [S1] K. which attenuates the low-frequency content of the uncertainty u x 1 s 2 .” We notice that the L1 reference system is equivalent to a disturbance observer type closed-loop system as shown in “Bridging Adaptive and Robust Control. however. “Disturbance rejection through an external model for non-minimum phase systems. The transfer function 12C 1 s 2 is a highpass filter. u ref 1 s 2 5 12C 1s 2 12C 1s 2 H 1s 2 (S4) proper C 1 s 2 . 1.” Trans. no.

it follows that sH 1s 2 is proper and exponentially stable. vc where v c . 0 is the filter bandwidth. we can derive uref 1 s 2 2u 1 s 2 5 C 1 s 2 1 2u 1 xref 1 s 2 2x 1 s 22 1 | h 1 s 22 5 2C 1 s 2 u 1 xref 1 s 2 2x 1 s 22 1 C 1 s 2 s 1 am | x 1s2. b (30) Subtracting (27) from (21) gives h1s2. in L1 adaptive-control architectures. Similarly. Notice that the sufficient condition for stability of the L1 reference system (24) ensures that (25) is meaningful. (29) 12G 1 s 2 u L1 12G 1 s 2 u L1 "G where | h1t2 ! | u 1 t 2 x 1 t 2 . Rewriting the plant dynamics (1) in the frequency domain. From (23) it follows that vc S ` vc S ` Therefore 7 G 1s 2 u 7 L1 5 7 1 12C 1 s 22 H 1 s 2 u 7 L1 # g 1 g 7 sH 1 s 2 u 7 L1 s 1 v c L1 1 5 7 sH 1 s 2 u 7 L1 . u 1 s 2 5 C 1 s 2 1 kgr 1 s 2 2u x 1 s 2 2h (26) 7 xref 2 x 7 L` # g 2u max C1s2 C1s2 g 7| g x 7 L` # g . that is. 5 H 1 s 2 1 C 1 s 2 kgr 1 s 2 1 1 12C 1 s 22 u x 1 s 2 2C 1 s 2 | 5 (27) This bound implies that the error between the states of the closed-loop system with the L1 adaptive controller and the L1 reference system. increasing the bandwidth of the fi lter leads to high-gain control with a reduced time-delay margin [3]. hence. using (20). (26). We further notice that this uniform » OCTOBER 2011 . For simplicity. the transfer function C 1 s 2 1 s 1 am 2 / b reduces to 1 s 1 am 2 / b. s 1 am (28) (31) Using the bound on 0 | x 1 t 2 0 in (16). However. and thus the L1 norms of both C 1 s 2 and C 1 s 2 1 s 1 am 2 / b are bounded. vc and hence the L1 -norm condition can be satisfi ed for a given plant (1) by choosing the fi lter C 1s 2 with suffi ciently large bandwidth. we obtain the uniform bound for the difference in the control signals given by 7 uref2u 7 L` # 7 C 1 s 2 u 7 L1 7 xref2x 7 L` 1 g C 1 s 2 s 1 am 2u max g . and. let vc C 1s2 5 . it follows that C 1 s 2 1 s 1 am 2 / b is proper and exponentially stable. the allowed bandwidth of the fi lter is limited by robustness considerations. as compared to its nonadaptive version. which uses the reference controller. Therefore. Therefore. s 1 vc s G 1s 2 5 1 12C 1 s 22 H 1 s 2 5 H 1s2 . Thus.Verifying the L1-Norm Bound T he stability of the L1 reference system is reduced to verifying the L1-norm condition in (24). the L1 controller has the same prediction error dynamics as in (11). can be uniformly bounded by a constant inversely proportional to the square root of the adaptation gain G. with C 1 s 2 5 1. This analysis illustrates the role of C 1 s 2 toward obtaining a uniform performance bound for the control signal of the L1 adaptive-control architecture. and (28). xref 1 s 2 2 x 1 s 2 5 G 1 s 2 u 1 xref 1 s 2 2 x 1 s 22 1 H 1 s 2 C 1 s 2 | which can be solved for xref 1 s 2 2x 1 s 2 to obtain xref 1 s 2 2x 1 s 2 5 C1s2 H1s2| h1s2. we obtain the bound 62 IEEE CONTROL SYSTEMS MAGAZINE Notice that without the lowpass filter. The question of the filter design to satisfy the condition in (24) is discussed in “Verifying the L1-Norm Bound. which is improper. we obtain x1s2 b 1 u 1 s2 1u x 1 s 22 1 xin 1 s 2 s 1 am h 1 s 22 1 xin 1 s 2 . it follows that | x 1s2 5 which leads to xref 1 s 2 2 x 1 s 2 5 C1s2 | x 1s2. b L1 "G Since the L1 adaptive controller uses the same state predictor (10) and adaptation law (12) as the MRAC with state predictor. 12G 1 s 2 u Because C 1 s 2 is strictly proper and exponentially stable.” Next. in the absence of the filter C 1 s 2 . we rewrite the control signal (18) as | 1 s 22 . and therefore 7 sH 1 s 2 u 7 L1 is bounded. s 1 vc Since H 1s 2 is strictly proper and exponentially stable. Thus. lim 7 G 1 s 2 u 7 L1 # lim 1 7 sH 1 s 2 u 7 L1 5 0. we cannot uniformly bound 0 uref 1 t 2 2u 1 t 2 0 as in (31). 12G 1 s 2 u b | h1s2.

` 2 . the performance and robustness of which can be analyzed using classical control techniques. and xm 1 t 2 [ R is the state of the ideal system (5). v gc It can be shown [5] that increasing G leads to higher gaincrossover frequency and thus reduced phase margin. Further analysis shows that increasing the adaptation gain. Consequently. corrupted with noise n 1 t 2 [ R. For this purpose. which improves disturbance rejection. the adaptation gain G in (35) resolves the performance tradeoff between disturbance rejection and noise attenuation. (12). (32) (33) ∧ σ −Γ s FIGURE 3 Closed-loop system with a model reference adaptivecontrol-type integral controller. a larger adaptation gain shifts the pair of the closedloop system poles closer to the imaginary axis. Figure 4(a) shows that the Nyquist plot of L1 1 s 2 does not cross the negative part of the real line. therefore. (38) s2 1 s 1 G where z 1 t 2 [ R is the measurement of the state x 1 t 2 [ R. Figure 3 shows that. consider the scalar plant # x 1 t 2 5 2x 1 t 2 1 u 1 t 2 1 s 1 t 2 . in the absence of input disturbance and measurement noise. for the performance analysis we consider the transfer functions from the input disturbance and the measurement noise to the plant control input and output. we conduct a preliminary robustness analysis of the L1 controller (18). on the other hand. x 1 0 2 5 x0. The block diagram of the closed-loop system is shown in Figure 3. we now focus on robustness and performance analysis of the MRAC architecture described in (6) and (8). Hence. In this case both the MRAC and L1 adaptive controller specialize to a linear model-following controller. whereas. This fact implies that Hun 1 s 2 acts similarly to a differentiator and thus leads to undesirable amplification of the measurement noise in the control channel. we assume that am 5 1 and b 5 1 in (1). the MRAC scheme (34)– (36) may develop high-frequency oscillations. similar to the tracking error. the closed-loop system response is identical to the response of the ideal system (36). The gaincrossover frequency v gc can be computed from 0 L1 1 jv gc 2 0 5 G v gc"v 2 gc 1 1 5 1. s 1 0 2 5 s0 . From Figure 3 we obtain which leads to the phase margin fm 5 p 1 /L1 1 jv gc 2 5 arctan a 1 b. H xs 1 s 2 5 s . z1t2 5 x1t2 1 n1t2. both performance bounds can be systematically reduced by increasing the rate of adaptation. (8) is equivalent to the MRAC (6). For the plant with input disturbance (32). For this purpose we consider the loop transfer function of the system in Figure 3 with negative feedback given by L1 1 s 2 5 G . on the one hand. The OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 63 . Thus. in the presence of fast adaptation. s1s 1 12 (39) # ^ 1 t 2 5 2Ge 1 t 2 5 2G 1 xm 1 t 2 2z 1 t 22 . and we analyze the performance of the closed-loop adaptive system in the presence of input disturbances and measurement noise. Next we investigate the robustness properties of MRAC. which in this case becomes # xm 1 t 2 5 2xm 1 t 2 1 r 1 t 2 . Next. s2 1 s 1 G Hun 1 s 2 5 2 G s11 . Model Reference Adaptive Control Having shown that the state predictor parameterization of MRAC (10). given by where s (34) We notice that all of the transfer functions in (37)–(38) have the same denominator. The adaptation gain G is located in the feedback loop of the control system. which gives a Hurwitz pair of closed-loop poles with the damping inversely proportional to "G and natural frequency proportional to "G. this architecture specializes to the integral controller ^ 1t2 1 r1t2. the loop gain and bandwidth of the closed-loop system are determined by G. reduces the gain of Hxs 1 s 2 at low frequencies. s 1s1G 2 Hxn 1 s 2 5 2 G . Hence. s 1s1G 2 (37) H us 1 s 2 5 2 G . u 1 t 2 5 2s ^ 1 t 2 [ R is an estimate of s 1 t 2 .bound is inversely proportional to the square root of the adaptation gain. the closedloop system has infinite gain margin gm 5 ` . s (35) G [ 1 0. xm 1 0 2 5 x0. and s 1 t 2 [ R is the unknown signal to be rejected by the control input u 1 t 2 [ R. (36) from which gain and phase margins can be computed. u r – σ 1 s +1 1 s +1 x xm n z – e Performance and Robustness Analysis Using a Simplified Scalar System Because a large adaptation rate in the case of MRAC leads to poor robustness characteristics. (8). Thus.

Hence. the effect of the lightly damped pole can be compensated or even canceled by an appropriate choice of C 1 s 2 . and transfer functions (43)– (44) reduce to (37)–(38). which is observed in MRAC.2 0 FIGURE 5 Closed-loop system with an L1 adaptive controller.8 −1 −1 C (s ) ∧ x – ~ x Γ s Fast Estimation Loop L1 Controller Γ = 10 Γ = 100 Γ = 1. for the L1 controller. s 1s1G 2 (44) reduction of phase margin with large G can also be observed in Figure 4(a). if increasing G improves the tracking performance for all t $ 0. 1. Hence. given by (12).2 0 Imaginary Axis ^ 1 t 2 2z 1 t 2 and G [ 1 0. Hence » OCTOBER 2011 . the closed-loop system recovers the ideal system in (36). that is.8 −0. similar to where | x 1t2 ! x ^ 1 t 2 for the con(18). However. then robustness degrades. in the absence of input disturbance and measurement noise. Hxn 1 s 2 5 2 2 s11 s 1s1G s 1s1G G C1s2. (43) a1 2 2 b . avoiding the undesired transient behavior.6 −0.4 −0. lightly damped poles arise in the estimation loop shown in Figure 5. ` 2 .8 −1 −1. On the other hand.0 −0. From Figure 5 we see that. transfer functions (43)–(44) containing this pole are followed by the lowpass filter C 1 s 2 . Thus.6 −0.4 Real Axis (a) −0. x (40) L1 Adaptive Control The parametric estimate. x ^ 1 0 2 5 x 0.8 −0. controller (42) reduces to the MRAC-type integral controller introduced in (34). which is implemented inside the controller block. including the transient phase. u 1 s 2 5 2C 1 s 2 s (42) −0. Plot (a) shows that the phase margin of the model reference adaptivecontrol-type integral controller vanishes as the adaptation gain G is increases. This fact implies that the L1 controller also results in lightly damped closed-loop poles in the presence of fast adaptation. while the bandwidth of the control loop is determined by the lowpass filter C 1 s 2 . 0.4 −0. we use a lowpass-filtered version of s trol law given by ^ 1s2 1 r1s2. The state predictor of the L1 adaptive controller. C 1 s 2 .2 Real Axis (b) 0 The block diagram of the closed-loop system is given in Figure 5. s 1 0 2 5 s0.2 −0. in the absence of the filter.2 Imaginary Axis r u σ n – 1 s +1 1 s +1 ∧ σ x z −0. s 1s1G 2 FIGURE 4 Nyquist plots of the loop transfer functions. is thus replaced by # ^ 1 t 2 5 2G | s x 1 t 2 . the adaptation rate G is the key to the tradeoff between performance and robustness in the design of MRAC. Notice that the fast. given by (10).000 −0. takes the form # ^ 1 t 2 5 2x ^ 1t2 1 u1t2 1 s ^ 1t2.6 −0. Moreover. plot (b) shows that the phase margin of the L1 adaptive controller approaches p / 2 as the adaptation gain G increases. Next. From the block diagram in Figure 5 we derive the transfer functions from the input disturbance and the measurement noise to the plant input and the output given by GC 1 s 2 1 G C 1 s 2 .000 −1 −0.2 Γ = 10 Γ = 100 Γ = 1.6 −0. This compensation allows for safe increase of the adaptation gain G without degrading the noise-attenuation properties of the system and without causing high-frequency oscillations in the control channel. similar to MRAC in Figure 3. The adaptation gain G affects only the fast estimation loop (red).4 −0. 64 IEEE CONTROL SYSTEMS MAGAZINE (41) Notice that the transfer functions (43)–(44) have the same denominator. H xs 1 s 2 5 H us 1 s 2 5 2 Hun 1 s 2 5 2G s11 C 1s2 .

In the foregoing analysis. This observation explains how the lowpass filter in L1 adaptive controller helps to decouple the estimation performance from the robustness of the adaptive controller. (48) (49) (47). which involves the plant. disturbances. However. The loop-transfer function of the system in Figure 5 is given by L2 1 s 2 5 GC 1 s 2 . and an unknown input gain. the gain-crossover frequency stays in the low-frequency range.000 100 Phase Margin (deg) 80 60 40 20 0 L1 Integral 0 200 400 600 Adaptive Gain (b) 800 1.02 dB with the increase of G. the limiting loop transfer function can be used to analyze the phase margin of the closed-loop adaptive system. as given in Figure 7(a) and (b) [62]. On the other hand. Consequently. The model reference adaptive-controltype integral controller has infinite gain margin while the phase margin diminishes as the adaptation gain increases. While the phase margin of the MRAC-type integral controller vanishes as the adaptation gain G is increased. We note that the dynamics of the adaptation loop in Figure 5 do not appear in the limiting loop transfer function « IEEE CONTROL SYSTEMS MAGAZINE 65 . In the direct architecture. s 1 vc (45) 15 Gain Margin (dB) 10 5 L1 0 0 200 400 600 Adaptive Gain (a) 800 1. Therefore. as G S ` . which consequently enables fast adaptation. ADAPTIVE CONTROL IN THE PRESENCE OF UNKNOWN INPUT GAIN In this section we revisit the main adaptive architectures for a class of multivariable systems with time-varying uncertainties.000 although similar results can be obtained using higher order filters. s 1 s 1 1 2 1 G 1 12C 1 s 22 (46) In the absence of the filter. while in the indirect architecture the plant parameters are estimated and used in the feedback law. The effect of the adaptive gain on the robustness margins of the two closed-loop systems. from Figure 3 we can see that for MRAC the lightly damped poles due to large adaptation gain are generated by the loop. this limiting loop transfer function cannot be used to analyze the gain margin of the closed-loop system with the L1 adaptive controller (40)–(42). We present in this section the indirect architecture using the state-predictor parameterization of OCTOBER 2011 Transfer function (47) has an infinite gain margin. the Nyquist plot in Figure 4(b) shows that the phase and the gain margins of the L1 controller are not significantly affected by large values of G. Figure 6 shows that the L1 adaptive controller has guaranteed phase and gain margins that are bounded away from zero in the presence of fast adaptation. and a phase margin of fm 5 p / 2. which in the case of the scalar plant (32) specializes to 1 1 u 1 s 2 1 s 1 s 22 . with C 1 s 2 5 1. the expression (46) leads to the limiting loop-transfer function L2 l 1 s 2 5 vc C1s2 . in the presence of fast adaptation. where the limiting loop transfer function (47) approximates the loop transfer function (46). Figure 6(a) shows that the gain margin of the L1 controller is finite and converges to gm 5 6. the controller parameters are directly updated. the looptransfer function (46) reduces to (39). Furthermore. even small plant uncertainty may cause these poles to drift to the right-hand side of the complex plane. that is. Although (46) has a more complex structure compared to (39).the compensation of these poles with the help of the filter occurs inside the controller and cannot be affected by the uncertainty or unmodeled dynamics possibly present in the plant. The transfer function (47) represents the loop transfer function of the L1 reference system. MRAC and L1. since the phase-crossover frequency tends to infinity as the adaptation gain G increases. We consider direct and indirect MRAC architectures. 5 s 12C 1 s 2 (47) FIGURE 6 Effect of high adaptation gain on the stability margins. that is. L2 1 s 2 5 L1 1 s 2 . xref 1 s 2 5 s 1 1 ref uref 1 s 2 5 2C 1 s 2 s 1 s 2 1 r 1 s 2 . However. These results are consistent with Figure 4. The (a) gain and (b) phase margins of the L1 adaptive controller are bounded away from zero. is presented in Figure 6. Then. causing closed-loop system instability. that is. we further consider the firstorder lowpass filter C1s2 5 vc . gm 5 ` .

MRAC. x 1 0 2 5 x0. d ⋅ ∧ ∧ kx = ΓProj(kx . b. use this architecture to develop the L1 adaptive controller. u 1 t 2 [ R is the control input. kr . –xe^Pb) ⋅ ∧ ∧ d = ΓProj(d. and. and (c) L1 adaptive controller.r ⋅ xm = Amxm + bkg r Ideal System ∧ ∧ ∧ u = kr r − kx x − d Control Law u ⋅ x = Amx + b (ω u + θ^x + σ ) Plant ∧ ∧ ∧ kx. The L1 adaptive-control architecture has the same state predictor and adaptation laws as the indirect MRAC architecture but a different control law. Am [ Rn 3 n is a known Hurwitz matrix that defines the desired dynamics for the closed-loop system. (b) indirect MRAC. –e^Pb) ⋅ ∧ ∧ kr = ΓProj(kr . ` 2 is the unknown plant . c [ Rn are known constant vectors. y 1 t 2 [ R is the regulated output. Thus. θ. –x Pb) Adaptation Laws (b) ∧ x x – ∼ x ∧ ∧ ∧ ω. 66 IEEE CONTROL SYSTEMS MAGAZINE (51) » OCTOBER 2011 where x 1 t 2 [ Rn is the measured state of the plant. and v [ 1 0. –ux ⋅ ∧ ∧ ∼^Pb) θ = Γ Proj(θ . –x Adaptation Laws (c) – ∧ ∧ ∧ ω. σ ∧ η u kD (s ) ∧ ∧ ∧ ∧ η = kgr – ω u – θ^x – σ ⋅ x = Amx + b (ω u + θ^x + σ ) Plant ∧ x ∼ x x r Control Law ⋅ ∧ ∧ ∧ ∧ ∧ x = Amx + b (ω u + θ^x + σ ) State Predictor ⋅ ∧ ∧ ∼^Pb) ω = ΓProj(ω. –xx ⋅ ∧ ∧ ∼^ σ = Γ Proj(σ . –xx^Pb) ⋅ ∧ ∧ ∼^ Pb) σ = ΓProj(σ . consider the plant dynamics given by # x 1 t 2 5 Amx 1 t 2 1 b 1 v u 1 t 2 1 u ^ 1 t 2 x 1 t 2 1 s 1 t 22 . s 1 t 2 [ R is the unknown disturbance. –ux ⋅ ∧ ∧ ∼ θ = ΓProj(θ . u 1 t 2 [ Rn is the vector of unknown timevarying parameters. The indirect MRAC architecture is based on a state predictor. similar to (18)–(19). θ. (50) y 1 t 2 5 c^ x 1 t 2 . σ FIGURE 7 Block diagrams of the adaptive-control architectures: (a) direct model reference adaptive control (MRAC). re^Pb) Adaptation Laws (a) x xm – e r ∧ ∧ 1 (k r − θ^x + σ ) u= ∧ g ω Control Law u ⋅ x = Amx + b (ω u + θ^x + σ ) Plant ⋅ ∧ ∧ ∧ ∧ ∧ x = Amx + b (ω u + θ^x + σ ) State Predictor ⋅ ∧ ∧ ∼^Pb) ω = Γ Proj(ω.

input gain. k ^ r 1 t 2 5 GProj 1 k ^ r102 5 k ^ . and kg ! 2 1 21 c^Am b (56) ensures that ym 1 t 2 tracks step reference inputs with zero steady-state error. Next we examine boundedness and asymptotic properties of the error dynamics (59) by considering the Lyapunov-function candidate | | | V 1 e 1 t 2 .2x 1 t 2 e^ 1 t 2 Pb 2R 1 Proj 1 k x | ^ 1 t 2 . assumes cancellation of uncertainties in the plant (50)–(51). | ^ 1 t 2 2 k 1 t 2 and | ^ 1 t 2 2 d 1 t 2 denote d 1t2 ! d where k x 1 t 2 ! k x x the parametric estimation error. ` . v g (57) for arbitrary Q 5 Q^ . d 1 t 2 . s 1 t 2 [ D ! 3 2sb. # 0 s 1 t 2 0 # ds. v min . 7 r 7 L` # r. and P 5 P^ . ` 2 is the adaptation gain. similar to (3). k x 1 t 2 . Using the property of the projection operator (S5). The unknown quantities are subject to the bounds v [ V ! 3 v min . (59) « IEEE CONTROL SYSTEMS MAGAZINE 67 . the projection bounds are set according to the available conservative bounds (52).d where G [ 1 0. and k r mate of kr ! kg / v in (57). xm 1 0 2 5 x0. We notice that the ideal system (54)–(55) is obtained from the plant (50)–(51) by applying the ideal controller uid 1 t 2 5 1 1 k r 1 t 2 2u^ 1 t 2 x 1 t 2 2s 1 t 22 . see “Projection Operator. v max 4 . k ^ 1t2. where U is a given convex compact set.k ^ repre^ . d ^ 1 t 2 by definition. d ^ 1 t 2 5 GProj 1 d ^1 0 2 5 d ^ . d 0 # (60) (61) (62) ym 1 t 2 5 c^xm 1 t 2 . 2 e^ 1 t 2 Pb 2 . # 7 u 1 t 2 7 # du . du. The tracking error signal e 1 t 2 ! xm 1 t 2 2 x 1 t 2 satisfies # | 1 2 1 2 | | e 1 t 2 5 Am e 1 t 2 1 v bk T x t x t 1 v bd 1 t 2 2 bv k r 1 t 2 r 1 t 2 . 2e^ 1 t 2 Pb 2R 1 2 d 1 t 2 Qe^ 1 t 2 Pv b 1 Proj 1 d | ^ 1 t 2 . e 1 0 2 5 0. d where k x x ^ 1 t 2 [ R is an estiis an estimate of d 1 t 2 ! s 1 t 2 / v . u1t2 5 k r x (58) ^1t2 [ R ^ 1 t 2 [ Rn is an estimate of k 1 t 2 ! u 1 t 2 / v . r 1 t 2 e^ 1 t 2 Pb 2 . of the adaptive estimates k x r The projection operator plays a crucial role in ensuring robustness and avoiding parameter drift. The control objective is to determine a state feedback controller u 1 t 2 such that y 1 t 2 follows a given bounded piecewise continuous input r 1 t 2 [ R. with the desired specifications given according to the ideal system # xm 1 t 2 5 Amxm 1 t 2 1 bkgr 1 t 2 . In the case of time-varying plant where k r 1 t 2 ! k r r parameters. For more details. (54) (55) | ^ 1 t 2 2k . the adaptation laws are given by ^ 1 t 2 . k x0 0 r0 sent the “best possible guess” of the ideal values of the unknown parameters. x102 5 x . k ^ 1 t 2 5 GProj 1 k ^ 102 5 k ^ . 2 x 1 t 2 e^ 1 t 2 Pb 2 . 0 . 0. r 1 t 2 e^ 1 t 2 Pb 2R 2 2 k r 1 t 2 Q e^ 1 t 2 Pv br 1 t 2 2 Proj 1 k r # 2 |^ # | 2 Q kx 1 t 2 kx 1 t 2 1 d 1 t 2 d 1 t 2R v G # 2 | | # # 2e^1 t 2 Qe 1 t 2 2 Q k ^ x 1 t 2 kx 1 t 2 1 d 1 t 2 d 1 t 2R v . which ensures boundedness ^ 1t2. we obtain a bound on the time-derivative of the Lyapunov function along the system trajectories (59)–(62) # # # V 1 t 2 5 e^ 1 t 2 Pe 1 t 2 1 e^ 1 t 2 Pe 1 t 2 # # # 1 |^ | | | | 1 t 2 k x 1 t 2 1 2| 1 1 2 kx d 1 t 2 d 1 t 2 1 2 k r 1 t 2 k r 1 t 2R v G |^ ^ ^ 1 5 e t 2 1 A^ mP 1 PAm 2 e 1 t 2 1 2e 1 t 2 Pv b 1 k x 1 t 2 x 1 t 2 # 2 |^ ^# | | ^ 1t2 1 t 2 kx 1 t 2 1 | 1 d 1 t 2 2 k r 1 t 2 r 1 t 22 1 Q kx d 1t2d G # 2 |^ # | # | 1 2^ 1 2 1 k r t k r t R v 2 Q kx 1 t 2 kx 1 t 2 1 d 1 t 2 d 1 t 2R v G | 1 2 ^1 2 5 2e^1 t 2 Qe 1 t 2 1 2 k ^ t Pv bx 1 t 2 x t Qe ^ 1 t 2 . u 1 t 2 [ U. and sb. the update laws for the parametric estimates use the projection operator. v max . k x x x x0 ^ 1 t 2 . Substituting (58) into (50) yields the closed-loop system dynamics | # |^ ^ 1t2r1t2. k r r0 # # ^ 1 t 2 . ds are known bounds. ^. x 1 t 2 5 1 Am 2 v bk x 1 t 22 x 1 t 2 2 v bd 1 t 2 1 v bk r 0 y 1 t 2 5 c^x 1 t 2 . k r 1 t 22 1 | | |2 |2 5 e^ 1 t 2 Pe 1 t 2 1 Q k ^ x 1 t 2 k x 1 t 2 1 d 1 t 2 1 k r 1 t 2R v .” Thus. 0 solves the algebraic Lyapunov equation A^ m P 1 PAm 5 2Q where xm 1 t 2 [ Rn is the state. (52) (53) for all t $ 0. sb 4 . G (63) which. G OCTOBER 2011 Direct MRAC The direct MRAC law uses the estimates of the unknown controller parameters in (57) and takes the form ^ 1t2r1t2 2 k ^ ^1t2x1t2 2 d ^1 t 2 .

y2 if f 1u2 . where f : R S R is the smooth convex function n Proj(θ . where lmin [ denotes the minimum eigen# value of the matrix. On the boundary f 1 u 2 5 1.y 2 2y 2 # 0. 0 # c # 1. In ^ the set 5 u [ Rn 0 0 # f 1 u 2 # 1 6 . the projection operator does not modify the vector. 0. y ) . The proofs of ultimate boundedness of the adaptive errors rely on (S5). y ) y ∇f (θ ) θ 0 f 1u2 ! 1 Pu 1 12 u ^u2u 2 max Pu u 2 max Projection Scaled by (1 − f (θ ) ) . the projection operator scales down the normal component of the vector as shown in (a). ||. and Pu .Projection Operator T he projection operator ensures boundedness of the parametric estimates by definition. if f 1 u 2 $ 0 and . if . the projection operator is defined as Proj 1 u. f y .f h . In the set between f 1 u 2 5 0 and f 1 u 2 5 1. 7 e 1 t 2 7 2 $ e2 max ! lmin 1 Q 2 Gv 2 min # then V 1 t 2 # 0. The definition of the projection operator proceeds by considering a compact convex set with a smooth boundary given by Vc ! 5 u [ R n 0 f 1 u2 # c 6 . where lmax [ denotes the maximum eigenvalue of the matrix. the normal component of the vector is canceled. if f 1u2 $ 0 and . FIGURE S3 Illustration of the projection operator. (a) f (θ ) = 0 f (θ ) = 1 y ∇f (θ ) Projection θ 0 Scaled by 0 Proj(θ . From (63) it follows that V 1 t 2 # 0 if V 1 t 2 $ Vmax. y i f 1 u2 . Inside the set bounded by f 1 u 2 5 0. # Notice that V 1 t 2 # 0. Vmax ! lmax 1 P 2 e2 max 1 v max Gv 2 min a 4u 2 max 1 4s2 b 1 k2 g v2 max where u max ! max 7 u 7 . if e 1 t 2 Qe 1 t 2 $ ^ 4v max Gv 2 min 1 u max du 1 sb ds 2 . if 4v max 1 u maxdu 1 sb ds 2 . The scaling factor is given by 1 2 f 1 u 2 .f ^y . and u [ V1 [64]. f ^y # 0.f .f || ||. As a result of this transformation. 0. e y. u max is the norm bound imposed on the vector u . the projection operator verifies the property 1 u2u * 2^ 1 Proj 1 u . the Proj 1 u . 0 is the projection tolerance bound of the choice. 0. y 2 operator subtracts a vector normal to the boundary V f 1u2 5 5 u [ Rn 0 f 1 u 2 5 f 1 u26 so that we obtain a smooth transformation from the original vector field y to an inward or tangent vector field for V1. (S5) f (θ ) = 0 f (θ ) = 1 (b) for given vectors y [ Rn. and » OCTOBER 2011 . u * [ V0 ( V1 ( Rn.y 2 ! y.y 2 does not alter y if u belongs to the set V0 ! 5 u [ Rn 0 f 1 u 2 # 0 6 . where 68 IEEE CONTROL SYSTEMS MAGAZINE v max Gv 2 min u md. u[U 3 1 v max 2 v min 2 2 b (64) 5 v max 4lmax 1 P 2 2 1 u max du 1sb ds 2 1 4u 2 a max 1 4sb lmin 1 Q 2 Gv 2 min 1 5 k2 g v2 max 1 v max 2 v min 2 2 b Therefore.f || Notice that the projection operator Proj 1 u . As shown in Figure S3.

k r 1 t 2 are uniformly bounded. y (66) ^ 1 t 2 [ Rn. Since x 1 t 2 5 xm 1 t 2 2 e 1 t 2 . s ^ 1 t 2 2 v. but we refer to the difference due to the predictor. | x 1 t 2 5 Am | x 1t2 1 b1v u ^1 t 2 x 1 t 2 1 s x 1 0 2 5 0. 1k r1t2 2 u ^ 1t2 g v (68) (72) Similar to the direct MRAC. Since ^ 1t2 2 | ^ 1 t 2 5 xm 1 t 2 . we obtain that V 1 t 2 # Vmax. (75) 14 lmin 1 Q 2 g0 "G . | | 1 t 22 5 | |2 1 t 2 |1t2 1 1 1v V1| x 1t2. We notice that the Lyapunov function (72) and its derivative (73) are independent of the choice of the control signal. s ^ 1 t 2 of the plant parameters to achieve the control u objective. We note that in the presence of timevarying plant parameters u 1 t 2 and s 1 t 2 asymptotic stability cannot be achieved. 1 4s2 b 1 kg v2 max Since V 1 0 2 # Vmax. The indirect MRAC control law uses the estimates v ^ 1t2. (74) and u max is defined in (64). and s ^ 1 t 2 [ R. and x of an exponentially stable ideal system. then x 1 t 2 is uniformly bounded. s x ^1 t 2 Pb 2 . ` 2 is the adaptation gain. The error in the indirect con^ 1 t 2 2x 1 t 2 . To derive the adaptation laws we consider a reparameterization of the direct MRAC architecture in Figure 7(a) using the state predictor # ^^ 1 t 2 x 1 t 2 1 s ^ 1 t 2 5 Am x ^ 1t2 1 b1v ^ 1t2u1t2 1 u ^ 1 0 2 5 x0 . and xm 1 t 2 is the state of an exponentially stable ideal system. This property is v achieved by the projection-based adaptation laws. and therefore. for all t $ 0. it follows that x 1 t 2 is uniformly bounded. v (69) # ^ ^ 1 t2 . x x (65) ^ ^ 1t2 5 c x ^ 1t2. the uniform bound for | x 1 t 2 (74) is independent of the control choice. System (65)–(66) replicates the plant with state x structure from (50)–(51). and therefore we do not distinguish them in the previous section as direct and indirect. the state predictor with the indirect control law (68) is equivalent to the ideal system in (54). | ^ 1t2 2 s1t2. v plant parameter estimates selected according to conservative knowledge of their ideal values. from this inequality we obtain the uniform bound 7| x 1t2 7 # where g0 ! u mi 2 2 . bound (74) implies that the OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 69 . (67) where | 1t2 ! v |1t2 ! s ^ 1t2 2 u 1t2. The control law is obtained by replacing the values of the unknown plant parameters in (57) by their estimates u1t2 5 1 ^^1 t 2 x 1 t 2 2 s ^ 1 t 22 . 2| ^ 102 5 s ^0 . 2x 1 t 2 | ^ 102 5 u ^0 . and ^ 0. Indirect MRAC The indirect MRAC scheme given in Figure 7(b) uses esti^ 1 t 2 [ Rn. v x ^ 1 t 2 Pb 2 . Å lmin 1 P 2 v 2 minG This definition of the control law requires that the estimate ^ 1 t 2 remain bounded away from zero. u mi ! 4u 2 max 1 4sb 1 1 v max 2 v min 2 Å lmin 1 P 2 lmax 1 P 2 1 u max du 1 sb ds 2 . 2u 1 t 2 | ^ 102 5 v ^ 0. s ^ 0 . where xm 1 t 2 is the state x1t2 5 x x 1 t 2 . lmin 1 P 2 7 e 1 t 2 7 2 # Vmax. In the case of known plant input gain v . for all t $ 0. x ^ 1 0 2 5 x0. u 1t2u 1t2 1 s x 2 1| V 1 t 2 # 2| x ^ 1 t 2 Q| G (73) | | | Hence. v u 1t2. s x ^ 1 t 2 Px G |2 1 t 22 1| u^1t2| u 1t2 1 s leads to # 2 ^ # # |1t2s 1 t 22 . d 1 t 2 . x Hence. ^ 1 t 2 5 G Proj 1 u u x 1 t 2 Pb 2 . Notice that substitution of (68) into (65) leads to the same ideal system as in (54) # ^ 1 t 2 5 Am x ^ 1 t 2 1 bk g r 1 t 2 . s (71) where G [ 1 0. Finally. for all t $ 0. u ^ 1 t 2 [ R of the plant mates v parameters but not the controller parameters. u ^ 0 are the initial conditions for the v max . which.u md ! 4lmax 1 P 2 1 u maxdu1sb ds 2 1 4u 2 max lmin 1 Q 2 2 1 v max2v min 2 2 . k x 1 t 2 . yields the bound 7e1t2 7 # v maxu md . with the unknown parameters replaced by their estimates. the signals e 1 t 2 . the indirect MRAC in Figure 7(a) and the direct MRAC in Figure 7(b) architectures are identical. v min. which trol scheme is the estimation error | x 1t2 ! x is governed by # | | 1t2u1t2 1 | | 1 t 22 . v u 1t2 ! u ^ 1t2. similar to (60)–(62) # ^ 1 t 2 5 G Proj 1 v ^ 1 t 2 . Hence. u (70) # ^ 1 t 2 5 G Proj 1 s ^ 1 t 2 . ^ 1 t 22 . The choice of the Lyapunov-function candidate | 1t2.

Substituting the reference controller from (81) into the plant dynamics (50) leads to the L1 reference model # xref 1 t 2 5 Am xref 1 t 2 1 b 1 v uref 1 t 2 1 u^1 t 2 xref 1 t 2 1 s 1 t 22 . Consider the plant (50). we ^ 1 t 2 . To derive the dynamics of the reference system for the plant (50)–(51). Thus. we obtain ^ ^1 t 2 x 1 t 2 2 s # ^ 1t 2 k gr 1 t 2 2 u ^ 1 t 2 5 Am x ^ 1 t 2 1 b av ^ 1 t 2 cf 1 t 2 * a x b ^ 1t2 v ^^1 t 2 x 1 t 2 1 s ^ 1t≤ 2. To show where the problem lies. v max 4 . and assume that (52) holds. dynamically by driving h L1 Reference System Similar to the scalar case with known plant input gain. k gr 1 t 2 2 v (79) Comparing (79) to the indirect MRAC control law (68). consider the filtered version of the control signal in (68) uf 1 s 2 5 Cf 1 s 2 u 1 s 2 . L1 Adaptive Control is exponentially stable and has unit dc gain C 1 0 2 5 1 [61]. for all v [ 3 v min. C1s2 5 v kD 1 s 2 1 1 v kD 1 s 2 (78) The L1 adaptive-control architecture for an unknown plant input gain is presented in Figure 7(c) [61]. Hence. the L1 controller has the same prediction error dynamics as the indirect MRAC (67). in steady state. The L1 adaptive controller uses the same state predictor and adaptation laws as the indirect MRAC given by (65)–(66) and (69)–(71). Then the controller (77) takes the form uref 1 s 2 5 kD 1 s 2 1 kgr 1 s 2 2 href 1 s 2 2 v uref 1 s 22 . may result in unpredictable consequences on the sysv tem’s performance. in the presence of an unknown plant input gain. Recall that. v (81) (80) where * denotes the convolution operator. The choice D 1 s 2 5 1 / s with k . 1u (76) ^ 1 t 2 may change rapNotice that the parameter estimate v idly due to the fast adaptation that. (82) » OCTOBER 2011 . the strictly proper transfer function 70 IEEE CONTROL SYSTEMS MAGAZINE where C 1 s 2 is defined in (78). The find that the L1 control law avoids division by v filter in the L1 control law (77) solves the design equation ^ 1 t 2 in (77) to zero [5]. v ^ 1 t 2 can be viewed as a timefore. Then uf 1 t 2 5 cf 1 t 2 * u 1 t 2 5 cf 1 t 2 * ¢ ^^1 t 2 x 1 t 2 2 s ^ 1t2 k gr 1 t 2 2 u ^ 1t2 v ≤. for the case of slowly varying parameters u 1 t 2 and s 1 t 2 . where href 1 s 2 is the Laplace transform of href 1 t 2 ! u^1 t 2 xref 1 t 2 1 s 1 t 2 . along with the fact that the convolution operator does not allow cancellation of ^ 1 t 2 . which compensates for ^ 1 t 2 in (76). notice that the requirement C 1 0 2 5 1 implies that D 1 s 2 must contain an integrator. The L1 adaptive controller instead generates the control u 1 t 2 from ^ 1s2. the lowpass filter cannot be directly applied to the control signal. the L1 adaptive controller considers a lowpass filter in the definition of the control law (18). 0 results in an exponentially stable strictly proper transfer function C1s2 5 kv . v To better understand the nature of the control law (77). respectively. the input to the ^ 1 t 2 < 0. s 1 kv ^ 1 s 2 does not multiply kD 1 s 2 . the closed-loop system with the L1 adaptive controller does not follow the ideal system (54)–(55) due to the limited bandwidth of the control channel enforced by C 1 s 2 . Let cf 1 t 2 be the impulse response of the transfer function Cf 1 s 2 . and the uniform bound on the prediction error given in (74) also holds. ThereNotice that in (77). consider the case of known plant parameters. the reference control law can be represented as uref 1 s 2 5 1 C 1 s 2 1 kgr 1 s 2 2 href 1 s 22 .tracking error can be arbitrarily reduced if we increase the adaptation gain G. Substituting this expression in the state predictor (65). Therefore. when u 1 t 2 is approximately a constant. and Cf 1 s 2 is a lowpass filter. control law (77) leads filter must be h to ^^1 t 2 x 1 t 2 2 s ^ 1t2u1t2 2 u ^ 1 t 2 < 0. in the case of known plant input gain. where u 1 s 2 is the Laplace transform of u 1 t 2 in (68). However. The key difference between the indirect MRAC and the L1 adaptive controller lies in the definition of the control law. u 1 s 2 5 kD 1 s 2 h (77) ^ 1 s 2 is the Laplace where k is a positive feedback gain. h transform of ^ ^1t2x1t2 2 s ^ 1t2u1t2 2 u ^ 1t2. xref 1 0 2 5 x0. ^ 1 t 2 ! k gr 1 t 2 2 v h and D 1 s 2 is a strictly proper transfer function such that. Therefore. in this approach v varying gain of the filter (78).

see “Special Form of State-toInput Stability. u[U Notice that the reference control signal (80) depends on the unknown parameters v . However. Setting the bandwidth of C 1 s 2 to be comparable with the control channel bandwidth. u 1 t 2 . For more details on the role of c0. which ensures that the transfer function c^ 0 H 1s 2 is minimum phase and has relative degree one. described in (26)–(31) and obtain [61] i xref 2 x i L` # i uref 2 u i L` # g1 "G g2 "G . Uniform Performance Bounds To derive the performance bounds. c^ 0H 1 s 2 (S7) kD 1s 2 c^ 0 xref 1s 2 . and c0 [ Rn is an arbitrary vector that makes c^ 0 H 1 s 2 minimum phase and relative degree one. we obtain the reference controller uref 1 s 2 5 kD 1s2 kg r 1s 2 2 which further can be rewritten as uref 1 s 2 5 kD 1 s 2 ^ ^ Qc 0 H 1s 2 kg r 1s 2 2c 0 xref 1s 2R .” OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 71 .Reference Controller for Multivariable Systems T he ideas discussed in “Bridging Adaptive and Robust Control” for the scalar case can be extended to the vector case. L we can rewrite the bound (84) as 7 xref 7 L` # 7 H 1 s 2 C 1 s 2 7 L1 kg 7 r 7 L 1 7 G 1 s 2 7 L1sb 1 7 xin 7 L 1 2 7G1s2 7 L L1 ` ` . (S6) kD 1s 2 c^ 0 xref 1s 2 5 kD 1 s 2 1 v uref 1 s 2 1 href 1 s 22 . we follow the steps similar to the scalar case. g0 is given by (75). compared to (80). (86) (87) where g1 ! g2 ! 12 7 G 1 s 2 7 L L 1 7 C 1 s 2 7 L1 g0 . the reference system in (81)–(82) achieves partial compensation of the uncertainty within the bandwidth of the control channel. see “Performance Bounds and Time-Delay Margin in the Presence of Nonzero Trajectory Initialization Errors.” The closed-loop L1 reference system can be written in the frequency domain as xref 1 s 2 5 H 1 s 2 C 1 s 2 kgr 1 s 2 1 H 1 s 2 1 12C 1 s 22 href 1 s 2 1 xin 1 s 2 . and xin 1 s 2 ! 1 sI2Am 2 21x0 is the Laplace transform of the ideal system response due to the initial conditions. Increasing the adaptation gain G is limited by the available CPU and highfrequency sensor noise. which also deviates from the ideal system. For details. In “Reference Controller for Multivariable Systems. Notice that these performance bounds are derived in the case where the initial conditions of state predictor (65)–(66) are equal to the initial conditions of plant (50)–(51). however. where L ! max 7 u 7 1. similar performance bounds can be derived for the case of their initialization mismatch. From (83) we obtain the bound 7 xref 7 L` # 7 H 1 s 2 C 1 s 2 7 L1 kg 7 r 7 L` 1 7 G 1 s 2 7 L1 1 L 7 xref 7 L` 1 sb 2 1 7 xin 7 L` . c^ 0 H 1s 2 Substituting this expression into (80) for the reference control signal and isolating u 1 s 2 . control law (S7) does not depend on the unknown plant parameters and the disturbance. Hence. i C 1 s 2 i L1 Lg1 1 v g ^ v c0 H 1 s 2 L1 (89) G 1 s 2 ! H 1 s 2 1 12C 1 s 22 . and the disturbance s 1 t 2 . Therefore. .” an alternative form of this controller is derived. Notice that (85) is consistent with the stability condition obtained for the scalar case (24) and can be satisfied for arbitrary H 1 s 2 by choosing C 1 s 2 with sufficiently large bandwidth. (85) Assuming 1 7 G 1 s 2 7 L1 . c^ 0 H 1s 2 ing (S6) by this transfer function. multiply- Control law (S7) produces the same control signal as (80). which facilitates implementation. (83) where H 1 s 2 ! 1 sI2Am 2 21b. In frequency domain (82) can be written as xref 1s 2 5 H 1s 2 1 v u ref 1s 2 1 href 1 s 22 . we consider the reference system (82) for zero initial conditions. by leading to the L1 reference system (81)–(82). (88) (84) 1 1 C 1 s 2 c^ 0 g g0 . we obtain We choose a constant vector c^ 0 satisfying the conditions of Lemma S2.” Notice that the performance bounds (86)– (87) are inversely proportional to "G. For more details about the bandwidth of the control system refer to “Available Bandwidth of Control Systems. The lowpass filter C 1 s 2 can thus be viewed as a means for resolving the ambiguity in (76) due to the convolution. . To this end. the ^ transfer function k D 1s 2 c^ 0 / 1 c0 H 1 s 2 2 is proper.

in particular. 2002. REFERENCE [S5] D.” IEEE Trans. standard MRAC architecture can be destabilized due to unmodeled dynamics at the plant input. (S8) where x 1t 2 [ Rn is the state. (S10) c^ o H 1s 2 where H1 1 s 2 ! C 1s 2 ^ co c^ o H 1s 2 LEMMA S2 If the pair 1 Am . letting n 1s 2 ! C 1s 2 u 1s 2 . is the matrix whose 1i. x 1 t 2 2 m for all t. pp. 9]. we can write C 1s 2 u 1s 2 5 C 1s 2 c^ o H 1s 2 u 1s 2 5 H1 1s 2 x 1s 2 . and b [ Rn. Am [ Rn 3 n. for x 1 0 2 5 0 and for all u 1s 2 we have mTx 1s2 5 mT N 1s 2 u 1s 2 5 0. j51 n has relative degree one with all its zeros in the open left-half plane.. Let H 1s 2 ! 1 s I2Am 2 21b. Consider an arbitrary vector c [ Rn such that c ^ 3s n21 c14 ^ ^ is a Hurwitz polynomial of degree n 2 1. it is possible to bound the norm of the output by a function of the norm of the input. It follows that. Lemma S2 implies that there exists co [ Rn ^ such that c^ o H 1s 2 has relative degree one. t [ 3 0. (S9) Controllability of 1 Am. x 1 02 5 0. t14 such that x 1 t1 2 5 xt1. this example demonstrates that. in the absence of specific modifications for bounding the adaptive parameters. The plant under consideration in [63] is the first-order exponentially stable system with unknown time-constant and dc gain and with two heavily damped.Special Form of State-to-Input Stability C onsider the system # x 1 t 2 5 Am x 1 t 2 1 bu 1 t 2 . Next. Indeed. b 2 implies that W has full rank. 422–436. The filtered input signal n 1 t 2 can thus be bounded in terms of the L` -norm of the system output. then for a given initial condition x 1 0 2 5 0 and arbitrary t1 and xt1. so that x 1s 2 5 H 1s 2 u 1s 2 . Sontag. 3. then there exists co [ Rn such that c^ o N 1s 2 /D 1s 2 has relative degree one. Liberzon. let H1s2 5 N 1s 2 . D 1s 2 . b 2 in (S8) is controllable. If W is not full rank. if the system is controllable. W must be full rank. whether it is possible to find a bound on the system input in terms of its output. “Output-input stability and minimum-phase nonlinear systems. Thus. we show that a bound can be derived for the lowpass-filtered input signal. and e-modification [35]. This conclusion contradicts the fact that x 1 t1 2 5 xt1 can be an arbitrary point in Rn. The following discussion highlights properties of the L1 adaptive controller that help to prevent this type of instability. and let co 5 1W 212 c . which is assumed to be bounded and piecewise-continuous. no. n3n n21 c ^ c^ 14 o W 3s is proper and exponentially stable. D. A. The opposite question can also be asked. namely. j 2 entry Wij is intro- BENCHMARK PROBLEM: ROHRS’S EXAMPLE In this section we analyze Rohrs’s example from [27] and [63]. 47. Specifically. While for system (S8) the answer to this question is negative [S5]. For this purpose. Automat. there exists u 1 t 2 . we have the bound 7 nt 7 L` # 7 H1 1 s 2 7 L1 7 xt 7 L`. vol. Then 21 b5 c^ o 1 s I2A 2 c ^ 3s n21 c14 ^ D 1s 2 D 1s 2 where D 1s 2 5 det 1 s I2A 2 . Thus. The norm of the system state for all t $ 0 satisfies the bound 7 xt 7 L` # 7 H 1 s 2 7 L1 7 ut 7 L`. for an exponentially stable proper linear system with zero initial conditions. let C 1s 2 be a strictly proper exponentially stable transfer function. Hence. Morse. such as projection [64]. then there exists nonzero m [ Rn such that mTN 1s 2 5 0. Contr. and E. u 1t 2 [ R is the control input. that is. S. and N 1s 2 is the n 3 1 vector whose i th component is the polynomial function of the form Ni 1s 2 5 a Wij s j21. high-frequency unmodeled poles given by » OCTOBER 2011 . D 1s 2 which implies that. The fact that arbitrarily small disturbances can 72 IEEE CONTROL SYSTEMS MAGAZINE lead to instability in adaptive-control schemes that do not have a protection for preventing parameter drift is illustrated in [65. s-modification [32]. deg 1D 1s 22 2deg 1 c^ 0 N 1 s22 5 1 and N 1s 2 has all its zeros in the open left-half plane. without invoking the derivatives of the output. and co H 1s 2 has all its zeros in the open left-half plane. which illustrates the robustness features of the MRAC architectures. u Next. Ex. Proof It follows from (S9) that for arbitrary co [ Rn c^ o H 1s 2 5 where W [ R duced in (S10).

the bound in (74) becomes zero initial conditions x | 1t 2 7 # r 1t 2 . subject to the bounds (52)–(53) on the uncertainties. This control system tions of the plant. lim r 1t 2 5 g0 "G . there exist the projection bound sb and the adaptation gain G such that the closed-loop system is bounded-input. where V 1 t 2 is the Lyapunov function (72) [4. For simplicity consider the plant (50)–(51) with constant uncertain parameters u 1 t 2 5 u. 1 ã lmin 1P 2 lmin 1P 2 G 2 2 un ! 4u 2 max 1 4sb 1 1 v max2v min 2 1 for all t $ 0. x results in the prediction dynamics (67) but with the nonzero | 102 5 x ^02x0. n -output exponentially stable proper transfer function F 1s 2 with impulse-response matrix f 1t 2 . Next. For an m -input. let T be the time-delay margin of the linear time-invariant system with the loop transfer function given by Lo 1s 2 5 and H 1s 2 ! 1s I2Am2b u^2 21b. this bound reduces to the one in (74).5]. tS` Using the bound on the prediction error in (S11) in the presence of nonzero initialization error. a lmax 1P 2 lmin 1Q 2 . 2. Then for the prediction error dynamics in (67) with non| 1 0 2 2 0. This result is semiglobal in the sense that for each bound 7 x0 7 # r0 on the initial conditions there exists a projection bound D 1 r0 2 that ensures that the closed-loop system has time-delay margin t. where | 1t 2 7 2.3] OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 73 . THEOREM S1 Consider plant (50)–(51) with constant uncertain parameters u 1 t 2 5 u. consider the state predictor (65)–(66) whose initial conditions differ from the initial conditions of the plant. we consider the same adinitial condition x aptation laws as in (69)–(71) and the L1 control law given by (77). We consider the state-predictor (65)–(66) with initial conditions that are different from the initial condi^ 1 02 5x ^0 2 x0. Next. we require the following notation. j 2 entry of the impulse-response matrix f 1t 2 . that is. (S11) a! 4sb ds . ^ v c0 H 1s 2 | 1t 2 is the inverse Laplace transform of x | 1s 2 ! and x in in ^ 02x0 2 . Further. arbitrary nonzero trajectory initialization errors lead to exponentially decaying transient errors in the input and output signals of the control system. T.Performance Bounds and Time-Delay Margin in the Presence of Nonzero Trajectory Initialization Errors n the design of state-feedback L1 controllers. and the adaptation laws (69)–(71). g3 1t 2 ! C H 2 1t 2 * 1 r 1t 2 1 7 x in H1 1s 2 5 C 1s 2 c^ C 1s 2 u ^ 0 . c. 7 uref 1t 2 2u 1t 2 7 ` # 1 | 1t 2 7 2 1 C 1t 2 g 1t 2 . for linear time-invariant plants. 12C 1s 2 C 1s 2 1 1 1 u^ H 1s 22 where fij 1t 2 is the 1 i. In the limiting case. let I 7 xref 1t 2 2x 1 t 2 7 ` # g3 1t 2 . C 1t 2 1 r 1t 2 1 7 x in H3 * 3 v H1 * which hold for all t $ 0.2. The next theorem [3] quantifies the time-delay margin at the plant input. To derive the performance bounds for the closed-loop adaptive system with the L1 adaptive controller. Furthermore. that is. then for each input timedelay t satisfying t . boundedoutput stable. x Let the adaptive controller be defined according to (77) with D 1s 2 5 1 / s. Lemma 2. H2 1s 2 5 1I2G 1 s 2 uT 2 21C 1s 2 . Thm. 1sI2Am 2 21 1 x 2 C F 1t 2 ! max a fij 1t 2 . If the L1-norm condition in (85) holds. while retaining the time-varying disturbance s 1 t 2 . Notice that this bound contains an exponentially decaying term and a constant component. H 3 1s 2 5 2 . when the exponential term diminishes. i 5 1.2. and all c0 [ Rn resulting in minimumphase and relative-degree-one c^ 0 H 1s 2 . n Å j 5 1 m The conclusion is that. 7x where un a V 102 2 b e2at un G r 1t 2 ! . the state predictor is initialized identically to the plant. we obtain the performance bounds [4. we analyze the degradation of the performance bounds due to initialization errors. that is. ^ 102 5x ^ 0 2 x0.

Therefore. k FIGURE 9 Rohrs’s example.85 sin 1 16. Usually. where the level of uncertainty is usually high.14. y 1 0 2 5 0. (94). Thus. The Bode plot given in Figure 8 illustrates this separation. overdamped unmodeled dynamics. k ^ r102 5 k k r0 2 .67 °. and k ^ 1 0 2 5 20.Magnitude (dB) 4 System Output 0 −20 −40 −60 0 −45 −90 −135 −180 −225 −270 10−1 Effect of Unmodeled Dynamics 2 0 −2 −4 0 y (t ) ym (t ) Phase (deg) 5 ∆ (s ) A (s ) A (s ) ∆ (s ) 100 101 Frequency (rad/s) 102 103 kr (t ) ∧ 10 Time (s) (a) 15 20 20 10 0 0 0 5 10 15 20 FIGURE 8 Impact of the unmodeled dynamics. which leads to instability and unbounded growth of the system output. The poles of the unmodeled dynamics are much faster than the pole of the plant. ^ r and k estimates k y point of view. ^ 1t2y1t2 1 k u1t2 5 k y r # ^ y0. The first reference input r1 1 t 2 5 0. ^ y1 0 2 5 k ^ k y 1 t 2 5 2e 1 t 2 y 1 t 2 . ym (t ) and (b) controller parameters ^ r 1t 2 . the unmodeled dynamics start significantly affecting the Bode diagram of the plant in the region where the magnitude of the plant’s frequency response is less than 220 dB. a properly designed control system does not attempt to act in the frequency range beyond the plant dynamics.3 1 2 sin 1 8t 2 (96) is also a sinusoidal reference signal but at a frequency that is approximately half of the phase-crossover frequency of (90)–(91). s13 where e 1 t 2 5 y 1 t 2 2 ym 1 t 2 . k # ^ . ^ 1 0 2 5 1. respectively. k r y ulation results from [63] are reproduced in figures 10 and » OCTOBER 2011 . s11 229 . Its phase-crossover frequency and the gain margin are v fc 5 16.09 rad /s and gm 5 24. The ^ are driven by the adaptation laws (93). ^ r 1 t 2 5 2e 1 t 2 r 1 t 2 . whose frequency is the phase-crossover frequency of the plant.1t 2 (95) Notice that the poles of the unmodeled dynamics are faster than poles of the plant and are separated from the plant dynamics in their frequency range. it is not difficult to handle unmodeled dynamics (91). Closed-loop model reference adaptivecontrol system with fast.70 rad/s and a phase margin of fm 5 107.3 1 1.62 dB. For simulations we consider the same reference inputs as in [63]. We use the same initial conditions as in [63]. D1s2 5 2 s 1 30s 1 229 y1s2 5 A1s2m1s2. We see that the unmodeled dynamics affect the plant’s Bode diagram in the region with less than 220 dB magnitude. m1s2 5 D1s2u1s2. k ^ y 1t 2 . Parameter Drift The MRAC controller for plant (90) takes the form ^ 1t2r1t2. where the input signals are attenuated. A1s2 5 (90) (91) ∧ (92) (93) (94) The plant has a gain-crossover frequency of v gc 5 1. (a) System output y (t ). while the second reference input r2 1 t 2 5 0. MRAC. from a classical control 74 IEEE CONTROL SYSTEMS MAGAZINE has the frequency equal to the phase-crossover frequency of the plant with the unmodeled dynamics (90)–(91). The corresponding feedback loop of the MRAC architecture is shown in Figure 9. k y (t ) −10 −20 0 5 10 Time (s) (b) 15 20 r kr ∧ u 229 s 2 + 30s + 229 Unmodeled Dynamics ky ∧ 2 s+1 Plant y FIGURE 10 Closed-loop model reference adaptive-control response to r1 1 t 2 .65. The adaptive estimates in (b) experience parameter drift. The control objective in [63] is given by the exponentially stable reference model ym 1 s 2 5 3 r1s2. The simnamely.

ym (t ) and (b) controller parameters k OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 75 . FIGURE 12 Closed-loop model reference adaptive control with the s-modification response to r1 1 t 2 . and we consider the same reference signal as in (95).8 0 −2 −4 0 5 10 Time (s) (a) 0 15 System Output 2 0. The control law remains unchanged and is given by (92).5 −1 0 10 20 30 Time (s) (b) 40 50 FIGURE 11 Closed-loop model reference adaptive-control (MRAC) response to r2 1 t 2 . k ^y 1t 2 . the simulation results for sm 5 0. (94) by adding damping terms.5 0 10 20 30 40 50 ky (t ) where sm [ 1 0.1. while tracking r1 1 t 2 . The damping term is set to sm 5 0.4 0. 0. the closed-loop system is unstable due to parameter drift.4 System Output 0.2 y (t ) 0 1 2 3 Time (s) (a) 4 ym (t ) 5 MRAC with the s-Modification The adaptation laws for MRAC with the s-modification are obtained from (93). ym (t ) shown in (a). However. ` 2 is a constant damping parameter. In Figure 11.6 0. which leads to instability and unbounded growth of the system output in (a). The parameter estimates and system output remain bounded. whose frequency is approximately half of the phase-crossover frequency of the plant. output y (t ). Figure 12 presents the simulation results for sm 5 0. and we do not observe parameter drift. The damping term is set to sm 5 0. k ^y 1t 2 . bursting occurs in the response of the closedloop adaptive system to the reference signal r2 1 t 2 . ^r 1 t 2 .1 in Figure 13 show that the parameter estimates experience drift. For the simulations we use the same parameters for MRAC as in the previous section. ym (t ) and (b) controller param^r 1t 2 .2 0 y (t ) −0. ^y 1 t 2 . This can be explained by the fact that ∧ −1 −1. Therefore.4 0. eters k System Output 11.5 0 0 10 20 30 40 50 k r (t ) ky (t ) ∧ 0 0 5 10 Time (s) (b) 15 20 ∧ –0.2 0 1 2 3 Time (s) (a) 4 ym (t ) 5 y (t ) ym (t ) 20 ky (t ) –5 1.6 0. the adaptive estimates in (b) experience parameter drift. (a) System ^r 1t 2 .2. k ^ r 1 t 2 5 2e 1 t 2 r 1 t 2 2sm k ^r102 5 k k r0 k r (t ) 3 ∧ (97) (98) 2 1 –0.5 0 10 20 30 Time (s) (b) 40 50 FIGURE 13 Closed-loop model reference adaptive control with the s-modification response to r1 1 t 2 .2.8 0. Next we consider the s-modification of the adaptation laws.2 0 −0. The value of sm is insufficient to ensure robustness to the given unmodeled dynamics. which recovers stability of the adaptive-control system. The parameter estimates for a short period of time move to the set for which the controller with frozen parameters is unstable. which leads to # ^ y0 . This behavior of the estimates results in temporary instability of the closed-loop MRAC and causes bursting of the system output y (t ). we can see that.5 ∧ kr (t ) 0 5 5 10 15 20 ∧ –10 1 0. In Figure 10. although the divergence rate is slower as compared to the results in Figure 10. In this case the parameter estimates converge. (a) System output y (t ). k ^ y 1 t 2 5 2e 1 t 2 y 1 t 2 2sm k ^ y1 0 2 5 k k # ^ .

G [ 1 0. s ^ 102 5 s ^ 0. 0 verifying 7 F 1s 2 7 L1 # LF. and where x the symmetric positive-definite matrix P 5 P^ . and known constants vl . u 0 # | ^ 1 t 2 5 G Proj 1 s ^ 1 t 2 . and F 1s 2 is an unknown exponentially stable transfer function whose dc gain has known sign. and thus the value of the damping term must be chosen large enough to ensure stability in the presence of unmodeled dynamics. du. s ^ 1t 2 [ R and ^ 1 t 2 [ Rn is the predictor state. Also assume that there exists known LF . u 0 5 u 0. and sb. 0 solves the the value of sm is not chosen sufficiently large to prevent the parameter drift. # x 1 t 2 5 23x 1 t 2 1 2 1 m 1 t 2 1 x 1 t 22 . 2 x t Pbx t . 2x 1 t 2 | ^ 102 5 u ^ . ` 2 is the adaptation gain. Next we present the elements of the L1 adaptive-control architecture for the class of systems with unmodeled dynamics at the plant input. x x ^ 1 t 2 5 c^x ^ 1t 2 . v x 1 t 2 u 1 t 22 . (S14) (S15) is strictly proper exponentially stable transfer function with dc gain C 1 0 2 5 1 for all F 1s 2 [ FD. x 1 0 2 5 x0. and r 1 t 2 5 3. vl # F 1 0 2 # v u. Figure 14 shows that. Finally. y1t2 5 x1t2. for all t $ 0. To analyze its performance we show simulations of the MRAC with the s-modification for step reference signals with different sizes. v ^ 102 5 v ^ 0. (S18) ^ 1 t 2 5 Am x ^ 1 t 2R . # # 7 u 1t 2 7 # du . Let u 1 t 2 and s0 1 t 2 . the system transient changes dramatically. Am [ Rn 3 n is a known Hurwitz matrix specifying the desired closed-loop dynamics. y 1t 2 [ R is the regulated system output. while v where x ^ u 1 t 2 [ Rn are the adaptive estimates driven by the adaptation laws. y (S19) ^ 1 t 2. ^ 1 t 2 5 G Proj 1 s ^ 1t 2 . 0 s0 1t 2 0 # ds . given by # |^1 t 2 Pbu 1 t 22 . we assume that we know a set FD of all admissible actuator dynamics. 2 x 1 t 22 . ` 2 are known bounds. s The L1 control law is given by ^ 1 s 2 2 kgr 1 s 22 . 0. u ^ 1 t 2 governed by with v # ^ 1 t 2 5 G Proj 1 v ^ 1 t 2 . without loss of generality. u 1 t 2 [ Rn is a vector of time-varying unknown parameters. s ^ 102 5 s ^ 0. r 1 t 2 5 2. 0 and a strictly proper. satisfy u 1 t 2 [ Q. The state predictor is given by # ^ ^1 t 2 x 1 t 2 1s ^ 1t 2 1 b Qv ^ 1t 2 u 1 t 2 1u ^ 1 02 5 x 0. ` . sb 4 . 2| ^ 102 5 v ^ 0. c [ Rn are known constant vectors. The performance of the closed-loop system is thus unpredictable for different reference commands. We notice that the goal of the s-modification is to recover stability. The design of the L1 adaptive controller proceeds by considering a positive feedback gain k . for the step reference commands r 1 t 2 5 1. Therefore. y 1t 2 5 c x 1t 2 . x ^ 1t2 5 x ^ 1t2. x 102 5 x0. ` . For this purpose. stable transfer function D 1 s 2 . 2 x v # ^ | ^ ^ ^ ^ 1 2 1 1 2 1 2 1 2 1 22 u t 5 GProj u t . the state predictor takes the form # ^ 1t2x1t2 1 s ^ 1 t 2 5 23x ^ 1t2 1 21v ^ 1 t 2 u 1 t 2 1u ^ 1 t 22 . # |^1 t 2 Pb 2 . where 229 u1s2. for design purposes. We perform tuning of the system for the case r 1 t 2 5 1. u x 1 t 22 . ^ 1 t 2 5 G Proj 1 v ^ 1t2. v # ^ 1 t 2 5 G Proj 1 u ^ 1 t 2 . where u 1 t 2 [ R is the control input. vu [ R satisfying 0 . ds [ 1 0. x ^ 1 0 2 5 x0 . s0 1t 2 [ D 5 3 2sb. ^ where. we assume F 1 0 2 . u 1 s 2 5 2kD 1 s 2 1 h (99) (100) (101) Next we apply the L1 adaptive controller from “L1 Adaptive Controller for Systems with Unmodeled Actuator Dynamics” to Rohrs’s example. which suggests the choice sm 5 0.L1 Adaptive Controller for Systems with Unmodeled Actuator Dynamics C onsider the system with multiplicative unmodeled dynamics located at the plant input [S6] given by (S12) (S13) # x 1t 2 5 Am x 1 t 2 1 b Qm 1t 2 1 u^1t 2 x 1 t 2 1 s0 1t 2R . such that C 1s 2 ! 1 1 kF 1s 2 D 1s 2 kF 1s 2 D 1s 2 (S17) where the state vector x 1 t 2 [ R n is assumed to be measured. b. we see that the damping term affects the robustness of the closed-loop system. y m1s2 5 ^ 1 t 2 and s ^ 1t2. (S16) | 1t 2 ! x ^ 1t 2 2x 1 t 2 . s2 1 30s 1 229 Then.2. and m 1 t 2 [ R is the output of the system m 1 s 2 5 F 1s 2 u 1 s 2 . we rewrite this example in state-space form as 76 IEEE CONTROL SYSTEMS MAGAZINE L1 Adaptive Controller » OCTOBER 2011 . 2 x s (S20) (S21) (S22) where Q is a known convex compact set. s0 1 t 2 [ R is a time-varying disturbance.

the open-loop system bandwidth ^ 1 t 2 adapt. The adaptive control law is given by G 1s 2 ! 112C 1s 22 H 1 s 2. r 1s 2 and h 1 2 forms of r t . pp. While the feedback ^y1t2 ^ 1 t 2 plays a similar role as the feedback gain k gain u ^ 1 t 2 appears in the in MRAC. The where h block diagram of the L1 adaptive-control system is given in Figure 15. Cao and N. New Orleans. 7x L` 0 1 ^ 1s 2 are the Laplace transwhere kg ! 21 / 1 c^A2 m b 2 . 5. |7 # g . u 1s 2 5 2kD 1s 2 1 h (S23) Then for all c0 [ Rn such that c^ 0 H 1 s 2 is minimum phase and relative degree one. and ^ 0 5 1. and thus avoid the parameter drift. 891–896. 10 4 . The ^ 0 5 0.5 4 . THEOREM 1] Assume that k and D 1 s 2 are chosen such that 7 G 1 s 2 7 L1L . Instead. 7 uref 2 u 7 L` # g2. 1. We see that the response both for plant output and the control signal scales by the value of the step reference command r 1 t 2 despite the nonlinear nature of the adaptation laws. and b [ 1 0. and G 5 1000. Next we simulate the L1 adaptive controller for step reference commands. The value of the constant u m depends on the known uncertainty bounds (S14)–(S16) and the adaptive control system parameters.14.5. h The L1 reference system for this architecture is given by # xref 1t 2 5 Am xref 1t 2 1 b Qmref 1t 2 1 u^1t 2 xref 1t 2 1 s0 1t 2R . and V 5 3 0. H 1s 2 ! 1s I2Am 2 21b. Figures 16–18 show that the L1 adaptive controller guarantees that both the plant output and the parameters remain bounded. r 1 t 2 5 2. as ^ 1 t 2 does in the L1 architecture. the adaptive parameter v feedforward path as a feedback gain around the integrator k / s and thus has the ability to adjust the bandwidth of the lowpass filter. For further insight into the L1 adaptive controller we analyze the implementation block diagram in Figure 15 from a classical control perspective. u initial conditions are set to x0 5 0. ` 2 is a constant depending on the plant uncertainty bounds. The expression for u m is given in [S6]. and href 1s 2 is the Laplace transform of href 1 t 2 ! u^1 t 2 xref 1 t 2 1 s0 1 t 2 . (S25) (S26) (S27) g1 ! where g0 ! 7 xref 2 x 7 L` # g1. The simulation plots for the inputs r1 1 t 2 and r2 1 t 2 . where (S28) C 1 s 2 c^ C 1s 2 0 g Lg1 1 g ^ g g0 . OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 77 . uref 1s 2 5 F1s2 yref 1t 2 5 c^xref 1t 2 . given by (95)–(96). ` 2 is an arbitrary constant. v ^ 0 5 0. “L1 adaptive controller for systems in the presence of unmodelled actuator dynamics. L ! max 7 u 7 1. The L1 adaptive controller has predictable performance in the presence of unmodeled multiplicative dynamics. um . xref 102 5 x0. which leads to ^ 1 t 2 and u changes as both v simultaneous adaptation on the loop gain and the phase of the closed-loop system. F 1s 2 c0 H 1s 2 L1 and u m [ 1 0. (S24) mref 1s 2 5 F 1 s 2 uref 1s 2 . IEEE Conf. The next theorem summarizes the main result for this adaptive architecture.65. Decision and Control. using the L1 adaptive controller with k 5 5. 2007. the output of which is the feedback signal of the closed-loop adaptive system. D 1 s 2 5 1 / s. ^ r 1 t 2 in (92) is Recall that in MRAC the feedforward gain k not engaged in the stabilization process and scales only the reference input. Note that the s -modification and the e -modification are used only to ensure boundedness of the parameter estimates. u[Q ^ 1 s 2 2 kgr 1 s 22 . REFERENCE [S6] C.” in Proc. with the L1 v adaptive controller. and ^^1 t 2 x 1 t 2 1 s ^ 1t 2 ! v ^ 1t 2 u 1 t 2 1 u ^ 1t 2 . The projection bounds are set to U 5 3 210. C1s2 1 kgr 1s 2 2 href 1s 22 . Å lmin 1P 2 G 12 7 G 1s 2 7 L L 1 7 C 1 s 2 7 L1 g0 1 b . s D 5 10. Hovakimyan. ^ 1t2x1t2 1 s ^ 1t2 ! v ^ 1t2u1t2 1 u ^ 1 t 2 and kg 5 3 / 2 .^ Lyapunov equation A^ m P 1 PAm 5 2Q for arbitrary Q 5 Q . 0. THEOREM S2 [S6. but do not affect the phase in the system. are given in figures 16–18. and r 1 t 2 5 3. LA. Figure 19 shows the simulation results for r 1 t 2 5 1. g2 ! g where xref 1t 2 [ Rn is the reference system state vector.

For a brief review of this control architecture. The estimates v driven by the adaptation laws (99)–(101). and s ^. where (104) u (t ) for r (t ) = 1 u (t ) for r (t ) = 2 u (t ) for r (t ) = 3 (b) FIGURE 14 Closed-loop model reference adaptive control with the s-modification response to r1 1 t 2 . ^ 1 t 2 5 2am x x (105) ^ 1 t 2 is the parameter estimate generated by the samwhere s pled-data update law ^ 1 t 2 5 2F21 1 Ts 2 m 1 Ts 2 | s x 1 iTs 2 . overdamped unmodeled dynamics. σ r – kg ∧ –k s u s2 + 229 30s + 229 2 s+1 Plant x Unmodeled Dynamics ω ∧ θ ∧ ^ . The adaptive controller must ensure that the output x 1 t 2 tracks the bounded reference signal r 1 t 2 with desired transient specifications given by the ideal system # xm 1 t 2 5 2amxm 1 t 2 1 bkgr 1 t 2 . 0 is a known plant parameter. (102) System Output 4 3 2 1 0 y (t ) 0 1 2 3 4 Time (s) (a) 5 r (t ) 6 10 5 where x 1 t 2 [ R is the measured state. see “ L1 Adaptive OutputFeedback Controller with Piecewise Constant Adaptation Law. t [ 3 iTs.5 tures of this adaptation law. Closed-loop L1 control system with fast. x 1 0 2 5 x0. b . Thus. (a) System output y(t ) (solid) for various step reference commands r(t ) (dashed) and (b) controller history u(t ). (106) In this section we present the L1 adaptive output feedback controller using piecewise constant adaptation laws.” We use a first-order system to elucidate the fea- PIECEWISE CONSTANT ADAPTATION LAWS FOR L1 CONTROLLER x 1t2 ! with Ts . and kg ! am/b.2. (103) u (t ) 0 –5 where am . 78 IEEE CONTROL SYSTEMS MAGAZINE » OCTOBER 2011 . The performance of the control system is different for different step reference commands. and x F 1 Ts 2 ! 3 e 2 am1Ts 2t2dt 5 c 0 Ts 1 2 am1Ts 2t2 Ts 1 e d ` 5 1 12e2amTs 2 . 0 is a known system parameter. applicable to non-SPR reference systems. xm 1 0 2 5 x0.u ^ are FIGURE 15 Rohrs’s example. Plant (102) can be written as 0 1 2 3 4 Time (s) 5 6 –10 # x 1 t 2 5 2amx 1 t 2 1 b 1 u 1 t 2 1 s 1 t 22 . b The state predictor for this example takes the form s1t2 ! # ^ 1 t 2 1 bu 1 t 2 1 s ^ 1t2. 1 i 1 1 2 Ts 2 . am2a x1t2 1 d1t2. a [ R is the unknown plant parameter with known conservative bounds. d 1 t 2 [ R is the input disturbance. The transient behavior of the system is unpredictable. 0 the adaptation sampling period. The damping term is set to sm 5 0. | ^ 1 t 2 2x 1 t 2 . let the plant dynamics be given by # x 1 t 2 5 2ax 1 t 2 1 b 1 u 1 t 2 1 d 1 t 22 . am a m 0 (107) (108) m 1 Ts 2 ! e 2 amTs.

we rewrite the plant (104) in frequency domain as x1s2 5 b 1 u 1 s 2 1 s 1 s 22 . a s s 1 am b (112) « IEEE CONTROL SYSTEMS MAGAZINE 79 . parameters v FIGURE 17 Closed-loop response of the L1 adaptive control system to r2 1 t 2 . The solution to (112) on the time interval t [ 3 t0. (a) System output y(t) . t0 1 Ts 4 for s 1 t 2 5 0 is given by 1 | ^ 1 t 2 1 1 2 e 2 am1t 2 t02 2 .4 0. we obtain the prediction error dynamics | x 1s2 5 5 b b 1 ^ 1 s 2b 2 1 u 1 s 2 1 s 1 s 22 au 1 s 2 1 s s 1 am b s 1 am b 1 ^ 1 s 2 2 s 1 s 2 b.2 0 1 2 3 Time (s) (a) System Output y (t ) ym (t ) 1 0. in reality. (113) Thus.0.6 0. x au 1 s 2 1 s s 1 am b (111) Subtracting (110) from (111). which is computed as the output of the lowpass filter C 1 s 2 with unit dc gain C 1 0 2 5 1. ym (t) and (b) controller ^ 1t 2 .5 System Output 0. the system output and parameter estimates remain bounded. (a) System output y(t) .. x 1t2 5 | x 0e 2 am1t 2 t02 1 s am 0 At the end of the sampling interval we have 1 | ^ 1 t0 2 1 12e2amTs 2 .2 0 −0. t0 1 Ts 4 . Therefore. OCTOBER 2011 where s 1 s 2 is the Laplace transform of s 1 t 2 . we set the disturbance during the next time interval t [ 3 t0. and the control system does not experience bursting. For this purpose. u ^ 1t 2 . s ^t. t0 2 for some t0 . u ^ t . which results in | x 1 t0 2 5 | x 0 2 0. The projection operator helps to avoid parameter drift. t0 1 Ts 4 to zero. The parameter estimates remain bounded.5 −1 0 1 2 y (t ) 3 ym (t ) 4 5 4 5 Time (s) (a) 5 10 5 0 2 0 –2 5 0 –5 ω (t ) ω (t ) ∧ ∧ 0 2 4 6 8 10 0 10 5 0 5 0 –5 0 2 4 6 8 10 θ (t ) 0 2 4 6 8 10 θ (t ) ∧ ∧ 0 2 4 6 8 10 σ (t ) 0 2 4 6 Time (s) (b) 8 10 σ (t ) ∧ ∧ 0 2 4 (b) 6 8 10 Time (s) FIGURE 16 Closed-loop response of the L1 adaptive control system to r1 1 t 2 .5 0 –0.8 1. To demonstrate the role of the sampling and the inversion of F 1 Ts 2 in the update law. u 1 s 2 5 C 1 s 2 a kr r 1 s 2 2 s b (109) ^ 1 s 2 are the Laplace transforms of r 1 t 2 and where r 1 s 2 and s ^ 1 2 s t . s 1 am (110) Assume that during t [ 3 0. is given by 1 ^ 1 s 2b . x 0 e2amTs 1 s x 1 t 0 1 Ts 2 5 | am Substituting the value of the adaptive estimate given by (106) yields 1 ^ 21 | x 1 t0 1 Ts 2 5 | F 1 Ts 2 m 1 Ts 2 | x 0 e2amTs 2 x 1 t0 2 1 12e2amTs 2 am 5| x 0 e2amTs 2 e2amTs | x 1 t0 2 5 0. ym (t) ^t . 0 the plant (104) is affected by the disturbance s 1 t 2 . respectively. However. The state predictor (105) can be similarly rewritten as b 1 ^ 1s2 5 ^ 1 s 2 b. we observe that the update law at each sampling time generates an estimate that perfectly compensates for the prediction error accumulated during the previous sampling period of the adaptation loop. and (b) controller parameters v The control signal. Next we check how the closed-loop system reacts to the accumulated prediction error | x 0. s ^ 1t 2 . the disturbance s 1 t 2 is not zero during t [ 3 t0.

which is not necessarily strictly positive real (SPR). y1 2 2f 1 t. the piecewise constant update ^ 1 t 2 is given by law for s | 1 iT 2 . Also. D "P Let (Am. ^ 1t 2 5 2F21 1 Ts 2 m 1 Ts 2 y s s s s (S37) Let the desired performance specifications be given by the exponentially stable minimum-phase transfer function M 1 s 2 of relative degree dr. s 1s 2 5 1 A 1s 2 2 M 1 s 22 u 1s 2 1 A 1s 2 d 1s 2 M 1s 2 . dr # nr is available. 0 4 ^ [ Rn. i 5 1. (S33) (S34) where P 5 P^ . y 1 t 2 [ R is the output. i 5 1. x ^ 1 02 5 0. 0. and y 1 iTs 2 . ^ 1t 2 5 c^x y (S35) (S36) substituting (S38) and (S39) in (S37) leads to the simplified form of adaptation ^ 1 t 2 5 2 1 e A mTs 2 I 2 2 1 AmeAmTsL21 11| y 1 iTs 2 . 2. Nowhere s tice that while s 1 t 2 [ R in (S33) is matched. so that Am is Hurwitz. 0 satisfies the algebraic Lyapunov equation A^ m P 1 PAm 5 2Q. the value of | x 1 t0 1 Ts 2 is usually not zero since the error dynamics are also affected by the additive disturbance. y 2 0 # L 0 y 0 1 L0. 1. 1. The state predictor is given by # ^ 1t 2 1 bu 1t 2 1 s ^ 1 t 2. y1. By setting the adaptation sampling time Ts small enough. 0 and L0 . with Ts F 1 Ts 2 ! L3 e 0 Am 1 Ts2t2 dt 5 1 e Am Ts 2I 2 . t [ 3 iTs. x 1 0 2 5 0. 1 i 1 1 2 T 2 . Namely. that is. The elements of L1 adaptive output-feedback controller are introduced next. c. which imposes hardware requirements. Notice that reducing Ts increases the values of the adaptation gains in F21 1Ts 2 .L1 Adaptive Output-Feedback Controller with Piecewise Constant Adaptation Law C onsider the single-input. where Q 5 Q^ . Then plant (S29) in terms of the ideal system can be rewritten as y 1s 2 5 M 1s 2 1 u 1 s 2 1 s 1 s 22 . thus making the error dynamics response equal to » OCTOBER 2011 . The adaptation laws for this L1 architecture are sampled in time. u 1 t 2 . 0 such that. D 1 c^ 1 "P 2 212^ 5 0. zero initial conditions. Here 11 ! 3 1. Letting y | the prediction error. s ^ 1t 2 [ Rn is the vector of adaptive parameters. y 1 t 22 . To further clarify the nature of this compensation. F21 1 Ts 2 cancels the effect of the error dynamics at fixed sampling times iTs. c be the sampled values of the prediction error. and D is a 1 n 2 1 2 3 n matrix that contains the null space of c^ 1 "P 2 21. The sampling period Ts is the fixed time-step during which ^ 1 t 2 take constant values during the adaptive parameters in s | 1t 2 ! y ^ 1 t 2 2y 1 t 2 be 3 iTs. 1 is higher than dimension of the scalar control input. single-output (SISO) plant y 1 s 2 5 A 1 s 2 1 u 1 s 2 1 d 1 s 22 . From (113) we see that F21 1 Ts 2 inverts the error dynamics in a discretized manner. x 1 0 2 5 0. that is. 1 i 1 12 Ts 2 for every i 5 0. the value of | x 1 t 2 can be kept small and arbitrary performance improvement can be achieved in the presence of disturbances. t [ 3 iT . b. 0 f 1 t. 2. We assume that there exist constants L . The proof of this result is given in [66]. We notice that F 1 Ts 2 is equal to the plant state at the time Ts for a constant unit 80 IEEE CONTROL SYSTEMS MAGAZINE control signal. and in the absence of the disturbance. y 1 t 2 5 c^x 1t 2 . and L! c c^ d. c. c. F 1 Ts 2 5 x 1 Ts 2 . y2 0 f 1 t. consider the definition of F 1 Ts 2 in (107). y2 2 0 # L 0 y12y2 0 . the uncertain- Therefore. (S29) where u 1 t 2 [ R is the control input. 3. 0. (S31) (S32) where F21 1 Ts 2 m 1 Ts 2 represents the update law gain. ^ 1 t 2 5 Am x x ^ 1t 2. for which only a known lower bound 1 . 1 i 11 2 Ts 2 . 0 is arbitrary. for all time t $ 0 and all y. (S38) (S39) m 1 Ts 2 ! Le AmTsL 2 111 . A 1 s 2 is a strictly proper unknown transfer function of unknown relative degree nr. s . (S30) ^ 1t 2 in (S35) is unmatched because its dimenty estimate s sion n . c ^) be a minimal realization of M 1 s 2 . Then the plant (S31) can be rewritten as # x 1t 2 5 Am x 1t 2 1 b 1u 1 t 2 1 s 1 t 22 . and d 1 s 2 is the Laplace transform of the time-varying uncertainty and disturbance d 1t 2 5 f 1 t.

and g2 1 Ts 2 are known class K functions of Ts. where g0 1 Ts 2 . where sref 1 s 2 ! 1 A 1s 2 2M 1 s 22 uref 1 s 2 1 A 1s 2 dref 1 s 2 M 1s 2 . which is the input that leads to the desired system response yid 1 s 2 5 M 1 s 2 r 1 s 2 by exactly canceling the uncertainty. and C 1 s 2 is a strictly proper system of relative degree dr. 0. 1. THEOREM 1] Assume that the lowpass filter C 1 s 2 for the given M 1 s 2 is chosen such that A 1s 2 M 1s 2 1 2 1 2 C s A s 1 1 12C 1 s 22 M 1 s 2 u 1 s 2 5 C 1 s 2 r 1s 2 2 c^ 1 sI 2 Am 2 21b where r 1s 2 is the Laplace transform of a bounded reference signal. The role of m 1 Ts 2 | x 1 iTs 2 is to compensate for the prediction error accumulated since the previous sample period of the adaptation loop. Next we define a filtered version of the ideal control law. Notice that the functions g0 1 Ts 2 . (S43) where G 1 s 2 ! H 1 s 2 1 12C 1 s 22 . (S40) c^ 1sI 2 Am 2 21s THEOREM S3 [66. Despite this fact. the adaption law (106) cannot be written for the case of continuous-time systems. Then the reference system in (S41)–(S42) is bounded-input. For details.The control law is given by C 1s 2 The next theorem summarizes the main result for this adaptive architecture. the quantity F 1 Ts 2 . and L is defined in (S30). where F 1 Ts 2 is a nonsingular matrix. with C 1 0 2 5 1. This fact implies that during the transient phase. uref 1 s 2 5 C 1 s 2 1 r 1s 2 2 sref 1 s 22 . given by (107). it is possible to achieve arbitrary improvement of tracking performance by uniformly reducing Ts. there exist a sampling time Ts such that 7| y 7 L` # g0 1Ts 2 . proportional to Ts. the control signal is continuous since the control signal is defined as the output of a lowpass filter.” where improvement of the transient performance is achieved by increasing the adaptation rate in the projection-based integral adaptation laws. for each Ts . g1 1 Ts 2 . Furthermore. nonstrictly positive real system can also be defined using standard projection-based adaptation laws. bounded-output stable. This fact implies that the inverse of F 1 Ts 2 exists for every Ts . we consider the ideal control signal uid 1 t 2 5 r 1 t 2 2 s 1 t 2 . as well between u 1 t 2 and uref 1 t 2 . We notice that the output-feedback L1 adaptive-control architecture for a minimum-phase. is not equal zero. including nonstrictly positive real systems. However. By using the ideas discussed in “Bridging Adaptive and Robust Control. OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 81 .” we can rewrite the equations for the L1 reference system in (S41)–(S42) to obtain the reference controller in the form uref 1s 2 5 C 1s 2 M 1 s 2 1 M 1 s 2 r 1s 2 2yref 1s 22 . 12C 1 s 2 This control law is similar to (S4). (S41) (S42) H 1s 2 ! is exponentially stable and satisfies the L1-norm condition 7 G 1 s 2 7 L1 L . g1 1 Ts 2 . The same holds also for systems of arbitrary dimension. yref 1 t 22 . The details on this adaptive architecture with stability and performance proofs can be found in [68]. see [66]. 0. 7 uref 2 u 7 L` # g2 1 Ts 2 . because F 1 Ts 2 approaches zero as Ts S 0. 7 yref 2 y 7 L` # g1 1 Ts 2 . These results are consistent with the results in the section “Adaptive Control in the Presence of Unknown Input Gain. ^ 1 s 2 . the tracking error between y 1 t 2 and yref 1t 2 . Similar to the state feedback case. and its structure allows the use of robust control for designing the filter for verification of the L1-norm stability condition in (S43) [S4]. Furthermore. as proved in [66] and [67]. is uniformly bounded by a constant and dref 1t 2 ! f 1 t. Thus. notice that. | x 11 i 1 1 2 Ts 2 5 | x 1 iTs 2 e2am Ts 2 m 1 Ts 2 | x 1 iTs 2 . This control law cancels the uncertainty within the bandwidth of the filter C 1 s 2 compatible with the control channel specifications and leads to the L1 reference system yref 1s 2 5 M 1s 2 1 uref 1 s 2 1 sref 1 s 22 . and g2 1 Ts 2 can be constructed based on the uncertainty bounds and adaptive control system parameters.

01 2 2 u 1 t 2 0. a 5 am.1s 1 1 1 1 12e2amTs 2 5 0. Figure 20 shows simulation results for the system without the plant uncertainty and the disturbance d 1 t 2 . if a 2 am. is possible if the plant pole in (102) is the same as the pole of the desired dynamics (103).99.4 System Output 4 2 y (t ) 3 2 1 0 0 1 2 3 Time (s) (a) 4 r (t ) u (t ) 0 –2 −4 5 0 1 2 3 Time (s) (a) 4 5 5 4 4 u (t ) u (t ) 2 0 –2 −4 0 1 2 3 Time (s) (b) 4 5 3 2 0 1 0 1 2 u (t ) for r (t ) = 1 u (t ) for r (t ) = 2 u (t ) for r (t ) = 3 3 4 Time (s) (b) 5 6 FIGURE 18 Control signal time-history for the L1 adaptive control system. the controller generates a parameter estimate.5. we consider the first-order plant (102). Otherwise. However. b 5 0. Illustrative Example To demonstrate the interpretation of the L1 adaptive controller with piecewise constant adaptation laws (105)–(109).5 and the same L1 controller.01 am Let the disturbance be given by d 1t2 5 100 1 u 1 t 2 0. that is.08 22 . while. as observed in Figure 20(c).07 2 2 u 1 t20. FIGURE 19 Closed-loop L1 adaptive controller response to several step reference commands. Figure 21 shows the simulation results for the unstable plant with the parameters a 5 20.04 2 2 u 1 t20. This error is treated by the L1 controller the same way as the error resulting from the disturbance. During other sampling periods both the prediction error and the parameter estimate remain zero. the plot of the prediction error given by (c) along with the plot for the parameter estimates given by (d) show that the L1 controller does not respond to the disturbance during the sampling period. during the second and the third intervals (114). Let the system parameters be given by a 5 am 5 b 5 1. which cancels the prediction error. While the controller tends to cancel the » OCTOBER 2011 . the disturbance is constant. The disturbance d 1 t 2 is zero everywhere except for three intervals of dura82 IEEE CONTROL SYSTEMS MAGAZINE tion Ts separated by intervals of duration 2Ts. the disturbance contains sine waves with different frequencies and bias. which results in additional prediction error.05 22 1 1 500 sin 1 1000t 2 25002 1 u 1 t20. (a) System output y (t ) (solid) for various step reference commands r (t ) (dashed) and (b) control history u (t ). 0. accumulated during the previous sampling period. (a) u(t) for r1(t ) and (b) u(t) for r2(t ). The control signal remains bounded and does not present high-frequency content. We set the sampling time Ts 5 0.01 s. the ideal plant dynamics (103) used by the inversion-based update law (106) are different from the true plant dynamics (102).02 22 1 1 300 1 500 sin 1 500t 22 1 u 1 t20. In Figure 20. (114) where u 1 t 2 denotes the unit step function. when the disturbance occurs. During the first interval. which results in F 1 Ts 2 5 and m 1 Ts 2 5 e2amTs 5 0. The perfect cancellation of the error caused by the disturbance d 1 t 2 in one sample step. The closed-loop response is scaled in accordance with the step reference commands. at the beginning of the following sampling period. For the control design we choose the first-order lowpass filter C1s2 5 1 .

02 0. choosing a smaller sampling period Ts helps to keep the prediction error small.06 Time (s) (a) 2.02 0.06 Time (s) (c) 0. a slightly modified version of the original two-cart system is used to illustrate the design methodology and performance OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 83 .1 0 0.04 0. control history u 1t 2 .06 Time (s) (d) 0.08 0.04 0.06 Time (s) (d) 0. The plots show that. (c) prediction error | 1t 2 .5 0 0 d (t ) 0.08 0.08 0. (b) | 1t 2 .5 –1 –1.1 30 20 10 0 –10 –20 –30 –40 u (t ) 0 0.06 Time (s) (c) 0. (c) prediction error x prediction error. (b) control history u 1t 2 .1 30 20 10 0 –10 –20 –30 –40 u (t ) 0 0.04 0. However. makes this error small. and (d) parameter estimate s ^ 1t 2 . The plant uncertainty prevents compensation of the prediction error during each sampling period.08 0. In the absence of the prediction error the controller generates a zero control input. at the beginning of the sampling period.06 Time (s) (b) 0.5 1 0 0.1 0 0.5 1 0.02 0. as predicted in [66] and [67].06 Time (s) (a) 5 4 3 2 1 0 –1 –2 –3 0 d (t ) 1000 500 0 –500 0.04 0. However.08 –1000 0.08 0. the controller cannot compensate for the prediction error completely due to the uncertain plant dynamics. Ts 5 0.04 0.5 2 1.02 0. the L1 adaptive controller generates a parameter estimate that cancels the prediction error accumulated during the previous sampling period. (a) State transient x(t). a small value of the sampling time.04 0.06 Time (s) (b) 0. (a) State transient x (t ) .1 FIGURE 20 System response to the disturbance d 1t 2 .1 150 100 50 0 –50 –100 –150 –200 –250 σ (t ) ~ x (t ) ∧ 0 0.4 x (t ) 2 4 x (t ) 2 0 –2 –2 0 0.1 300 200 100 0 –100 –200 –300 –400 –500 σ (t ) ~ x (t ) ∧ 0 0.02 0.02 0.02 0. and (d) parameter estimate s ^ 1t 2 .5 0 –0.08 0.02 0.01 s. namely. x 1000 x (t ) 2 4 x (t ) 2 0 –2 –2 0 0. BENCHMARK PROBLEM: TWO CARTS The two-cart mass-spring-damper (MSD) example is a benchmark problem for robust control design.08 –1000 0.1 FIGURE 21 System response to the disturbance d 1t 2 . In [69].04 0.5 1 0 0.04 0.

u (N) 0. and u 1 t 2 is modeled as white noise. and u 1t 2 is the control force. The signal d 1t 2 is a random colored disturbance force. and k2 are known. where the state vector is # # x^ 1 t 2 5 3 x1 1 t 2 x2 1 t 2 x1 1 t 2 x2 1 t 2 d 1 t 24 ^ and 0 0 k1 A 5 E2m 1 k1 m2 while u 1 t 2 is an additive sensor noise affecting the single measurement. defined by E 5 u 1 t 26 5 0. The system output tracks the reference commands with the desired control specifications.” Additional explanations and simulations can be found in [6]. 0 0 k1 m1 1 k2 2k1 m 2 1 0 b1 2m 1 b2 m2 0 1 b1 m1 1 b2 2b1 m 2 0 0 0 U. whose masses are m1 and m2.15 0. respectively. Only the spring stiffness coefficient k1 is unknown. s1a FIGURE 23 Closed-loop response to a step input of 1 m with k1 5 0. E 5 u 1 t 2 u 1 t 26 5 1026d 1 t 2 t 2 . Only x2 1 t 2 is measured.2 1 0.6 0. (a) System output x 2 (t ) and (b) control history u (t ). are given by m1 5 m2 5 1 kg.2 0 –0.2 0. 0 B^ 5 S 0 0 1 m1 0 0 0T. The parameters in (115)–(117). with zero-mean and unit intensity. that is. respectively.25 # k1 # 1. however. The states x1 1 t 2 and x2 1 t 2 represent the positions of the two carts. (117) The state-space representation is given by # x 1 t 2 5 Ax 1 t 2 1 Bu 1 t 2 1 Lj 1 t 2 .1. Next. The block with e2st represents the delay time t. we revisit the two-cart example with the L1 output feedback adaptive controller presented in “L1 Adaptive OutputFeedback Controller with Piecewise Constant Adaptation Law. The control signal remains bounded and does not exhibit high-frequency oscillations. b2. a 5 0.8 r (t ) x2(t ) m2 x2 (m) b2 0. and d1s2 5 a j 1 s 2 . The carts are mechanically connected between each other and the rigid surface by means of spring-damper couples k1– b1 and k2– b2.x1(t ) k1 u (t ) e−sτ m1 b1 x2(t ) d (t ) k2 1. 0. J 5 1. a . The masses of the carts and the parameters b1. respectively. Only x2 1 t 2 is measured.75. and u 1 t 2 is the control force. y 1 t 2 5 Cx 1 t 2 1 u 1 t 2 .1 N-s / m. The disturbance force d 1 t 2 is modeled as a first-order colored stochastic process generated by driving a lowpass filter with continuous-time white noise j 1 s 2 . » OCTOBER 2011 (116) The control objective is to design a control law u 1 t 2 so that the mass m2 tracks the reference step signal r 1 t 2 following 84 IEEE CONTROL SYSTEMS MAGAZINE . b1 5 b2 5 0.05 0 –0.05 0 100 200 300 Time (s) (b) 400 500 of the robust multiple model adaptive control. lower and upper bounds for k1 are known. independent of j 1 t 2 . which acts on the mass m1.25 and t 5 0.4 0. k2 5 0.2 0 100 200 300 Time (s) (a) 400 500 FIGURE 22 The two-cart mass-spring-damper system.05 s. 2a 1 m2 (115) while the spring constant k1 is unknown with known bounds given by 0.15 N / m.1 0. The signal d 1 t 2 is a random colored disturbance force acting on the mass m2. 0 0 L^ 5 S 0 0 0 0 a T . The two-cart system is shown in Figure 22. The states x1 1 t 2 and x2 1 t 2 represent the absolute position of the two carts. which are constant and known. whose masses are m1 and m2. C 5 S0 1 0 0 0T .

for additional details see [7] and [8].0 s2 1 0. the control signal remains bounded and does not exhibit high-frequency oscillations. Figures 23 and 24 show the response of the closed-loop system with the L1 adaptive output feedback controller to step reference inputs of different amplitudes and for different values of the unknown parameters k1 and t. As part of this project.05 s. such as changes due OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 85 . whose maximum possible value is 0. while minimizing the effects of both the disturbance and the sensor noise present in the system. τ = 0. τ = 0. which makes the amount of force exerted through the spring to the mass m2 uncertain. adaptive control has long been seen as an appealing technology with the potential to reduce loss-of-control (LOC) accidents and improve aircraft performance by reducing pilot workload during challenging flight conditions and in the event of severe failures or vehicle damage.05 0 0 –0.1. designed to provide a flexible research environment with the ability to conduct rapid prototyping and testing for control algorithms in extremely adverse flight conditions. one of whose primary objectives is to advance and transition adaptive flight control technologies as a means of increasing safety.05} x2(t ) {k1 = 0. s 1 2.66s 1 0. For application of the L1 adaptive output feedback controller to the two-cart example.05} x2(t ) {k1 = 0. while the sample time for the adaptation loop is set to Ts 5 1 ms.75.15 x2 (m) 0.02} x2(t ) {k1 = 1. τ = 0. Taking into account the fact that only x2 is measurable.25.05 0 100 200 300 Time (s) 400 500 r (t ) x2(t ) {k1 = 1.2 1 0. an L1 adaptivecontrol law was flown in a series of remotely piloted flight tests on the NASA AirSTAR Generic Transport Model (GTM) aircraft.19 . The control signal changes with different values of the parameters. see “NASA’s AirSTAR. τ = 0. τ = 0. Moreover the model contains the unknown time delay t in the control channel.19 5 4 SAFETY-CRITICAL FLIGHT CONTROL APPLICATION During 2007–2010. τ = 0.05} u (t ) {k1 = 0. For further details on the program and the GTM aircraft. τ = 0.8 s 1 3. the design procedure described in [70] leads to M1s2 5 C 1s 2 5 1 .6 0. s 3 1 1.2 –0.17s 1 0. while minimizing the effects of the disturbance d 1 t 2 and the sensor noise u 1 t 2 . the closed-loop system has similar response. such as inflight failure emulation and flight at high risk upset conditions.8 0.2 0. a desired model.35. For all cases of the unknown parameters.04} (b) FIGURE 24 Closed-loop response to a 1-m step input for various values of the unknown parameters k1 and t. the NASA Aviation Safety Program created the Integrated Resilient Aircraft Control (IRAC) Project.25.(a) System output x2 (t ) and (b) control signal u (t ).2 u (N) 0 100 200 300 Time (s) 400 500 0. The control problem is challenging because the control signal u 1 t 2 is applied through the uncertain spring.4 s 2 1 0. Flight Control Law Design The flight control system (FCS) developed for the GTM aircraft has as its primary objective achieving tracking for a variety of tasks with guaranteed stability and robustness in the presence of uncertain dynamics.05} u (t ) {k1 = 0. As we can see. In fact.35.18s 1 0. a twin-turbine-powered and dynamically scaled model of a civil transport aircraft.4 0.35.052 0.75.1 0.3 s2 1 2.04} (a) u (t ) {k1 = 1. the L1 adaptive controller drives the mass m2 to the desired position in about 15 s for arbitrary values of the unknown parameters within given bounds. τ = 0.02} u (t ) {k1 = 1.” This section provides an overview of the results from these flight tests. we have a noncolocated actuator problem [69]. however.35.

which can make recovery to normal flight difficult. weighs 54 lb at takeoff. The AirSTAR is a stateof-the-art facility designed for the purpose of investigating and validating high-payoff technologies aimed at the LOC problem using subscale flight vehicles without excessive risk [S7]. see “Aircraft Handling Qualities and Their Specifications. and sideslip-angle b adaptive command augmentation system (CAS). and has a flight time of approximately 18 min.5-ft wingspan. AirSTAR vehicles are designed for testing in regimes beyond the normal flight envelope or with degraded control characteristics. see Figure 1. LOC accidents are of great concern because they are usually catastrophic. twin-turbine-powered generic transport model. The aircraft is outfitted with full flight test instrumentation. These accidents are complex in that they typically have many causal factors and precursors.5% dynamically-scaled civil transport aircraft. one essential objective for safe flight under adverse conditions is for the aircraft to remain within the extended flight envelope since no guarantees for recovery can be made once outside the boundary as discussed in this article. and a 6DOF INS/GPS package including rate gyros and accelerometers. twin-turbine-powered. such as hardware failures or system faults. This requirement implies that the control law action in the initial 2–3 s after initiation of an adverse condition is the key to safe flight.” In particular. the L1 adaptive controller provides command-tracking capabilities in both » OCTOBER 2011 . while potentiometers mounted on the rotation axes of all control surfaces measure control-surface deflections. For example. system identification. Some of the key research areas are adaptive flight control technologies. data mining. controlsurface faults. [72] under nominal flight conditions with graceful degradation under significant adversity. The AirSTAR infrastructure is comprised of the research flight vehicle shown in Figure S4 and the mobile operation station (MOS). LOC accidents are known to exhibit post-stall angles of attack or steep pitch and bank attitudes. Consequently. while uplink commands are transmitted at 200 Hz. and they are difficult to analyze because often involve excursions beyond the normal flight envelope. The L1 control system used for this application consists of a nonadaptive stability augmentation system (SAS) for pitch and roll. roll-rate p. For more information on handling qualities.5% dynamically scaled.) to rapidly varying flight conditions during standard maneuvers and unexpected failures. The model has a 6. (Photos courtesy of Irene Gregory/NASA Langley Sean Smith. as well as a three-axis angle-of-attack a. [S8]. The test aircraft is a 5. and 5. flight dynamics modeling. Downlink data update rates vary from 5 Hz on the GPS data to 200 Hz on the data from analog sensors. which is based on the theory presented in “ L1 Adaptive Controller for Systems with Unmatched Uncertainties. such as unusual attitudes and post-stall flight. static and dynamic pressure sensors. structural damage such as missing wing segments. Dynamic scaling allows subscale flight test results to be applied to full-scale aircraft. and sensor failures. The primary AirSTAR flight test vehicle. Engine speed is also measured through an engine control unit. involving loss of aircraft and lives. is a remotely piloted. Flight validation of control technologies aimed at the LOC problem is recognized as a significant challenge due to the difficulties and risks associated with full-scale testing of transport airplanes in abnormal flight regimes. and fault-detection methods.” Hence. the Generic Transport Model (GTM) tail number T2. The GTM T2 model has a (a) (b) FIGURE S4 AirSTAR flight test vehicle with tail number T2 performing (a) takeoff and (b) landing. the adaptive controller 86 IEEE CONTROL SYSTEMS MAGAZINE is expected to learn fast enough to keep the aircraft within the extended flight envelope. All of these requirements are expected to be achieved while providing Level 1 handling qualities [71]. These conditions include upsets.NASA’s AirSTAR oss of control (LOC) is one of the leading contributors to the fatal accident rate of large commercial transport airplanes. including angle-ofattack and angle-of-sideslip vanes. L The AirSTAR infrastructure is designed to support aeronautics research under adverse conditions.

AIAA Guidance. namely.. may appear as an unmatched component in the inner-loop aircraft dynamics being controlled. Navigation and Control Conf.5-ft wingspan. Limited test volume combined with the high airspeeds of the test aircraft add to the pilot’s workload. “A flight control system architecture for the NASA AirSTAR flight test infrastructure. E. The MOS is an integrated ground station and control room and comprises a research pilot station as well as systems and engineering stations. The vehicles and operational concept used by AirSTAR present unique challenges. Foster. and has a flight time of approximately 18 min. the primary purpose of AirSTAR is to perform flight research and flight validation of multiple research control laws in adverse flight conditions. the AirSTAR concept of operations distinguishes three sequential phases. Downlink data update rates vary from 5 Hz on the GPS data to 200 Hz on the data from analog sensors.” in Proc. Honolulu.6. and landing. while uplink commands are transmitted at 200 Hz. A. Head-up and head-down synthetic vision displays drawn from aircraft sensor data and a local terrain database support the research pilot during this phase. reduced control-surface activity. “Practical application of a subscale transport aircraft for flight research in control upset and failure conditions. airdata probes mounted on each wingtip measure angle of attack.” in Proc. AIAA-2008-6200. extending to approximately 2000 ft in altitude. a reversionary flight control capability is provided by a safety pilot. estimated attitude angles. which is essential for achieving desired performance. these dynamics may have time scales comparable to the time scale of the pilot dynamics as well as the dynamics that the pilot is trying to control. In the research flight phase. reducing a pilot’s workload. Murch. and static pressure. thus reducing available pitch-control authority by 50% and destabilizing the aircraft by a prescribed amount. and experiment reproduction. and graceful degradation OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 87 . Third. a research pilot executes flight test maneuvers from a research cockpit located in the pilot station inside the MOS. [S8] A. Computers in the MOS execute research flight control laws and interface with the research pilot station using a direct connection by cable. HI. Aircraft dynamic failures are emulated using software to schedule appropriate actuators to provide desired degradation. and improving the aircraft’s handling qualities. REFERENCES [S7] K.. Actuator commands from research pilot inputs and the ground-based flight control computer are transmitted to the aircraft by the telemetry uplink. which presents unique challenges to flight testing in an efficient and safe manner. AIAA Guidance. For each flight. Second. angular-rate measurements. the aircraft must remain within visual range of the safety pilot. 2008. and A. While the model is being controlled from the research pilot station. The total downlink and uplink closed-loop data latency is estimated to be below 25 ms. such as airspeed and bank angle. Moreover. the effects of unregulated variables. unmodeled nonlinearities. nominal and off-nominal conditions without the presence of a nonadaptive CAS. which limits the test volume to an approximately 1 / 2-mi radius circle around the safety pilot. For example. M. cross-coupling between axes. First and foremost. then its undesirable effects in the inner-loop dynamics may require excessive pilot compensation or may lead to adverse aircraft-pilot interactions. If a problem were to arise. Morelli. The L1 adaptive FCS is designed to satisfy the same requirements as any standard classical flight control system. Murch. dynamically scaling an aircraft results in increased model airspeed. The MOS provides also a real-time simulation environment for pilot training. In fact. Additionally. as well as increased pilot workload relative to a model that is only geometrically scaled. For subscale platforms. The aircraft is outfitted with a 6DOF INS/GPS package. spoilers scheduled with roll rate are used to degrade lateral stability. 2008. if the design of the inner-loop FCS does not account for this uncertainty. This design provides a flexible environment to conduct flight research and evaluate research flight control laws with minimal risk. control law testing. J. research flight. and dynamic asymmetries. Engine speed is also measured through an engine control unit. M. angle of sideslip. Cunningham. the adaptive controller provides automatic compensation for the undesirable effects of unmatched uncertainty on the output of the system. weighs 54 lb at takeoff. The MOS is connected to the aircraft through an L-band with 1–2 GHz telemetry uplink and an S-band with 2–4 GHz telemetry downlink and provides realtime telemetry processing and monitoring capabilities. good flying qualities. HI. the safety pilot can override the research pilot control inputs and fly using a hobbyist radio-control interface. while potentiometers mounted on the rotation axes of all control surfaces measure control-surface deflections. takeoff. These factors make a dynamically scaled model more challenging to fly and less forgiving of mistakes. AIAA-2008-6990. under the current AirSTAR concept of operations. Further details on aircraft flight dynamics can be found in [73]. which provides three-axis linear accelerometer measurements. wind loading. dynamic pressure. Honolulu. Therefore. Similarly. while keeping the aircraft within visual range of the safety pilot. V. namely. Takeoff and landing are accomplished by a safety pilot using direct visual contact and conventional radio control equipment. such as pilot-induced oscillations. longitudinal stability degradation is achieved by commandeering two of the four elevator segments from the control law path and scheduling these segments with angle of attack to adversely change the pitching moment. Navigation and Control Conf. angular rates. and GPS position and velocity. precision tracking.

Further details about the design and robustness properties of the L1 adaptive flight control law for the GTM can be found in [74] and [75]. one with a feedback stability augmentation system. Moreover. REFERENCE [S9] J. Specifically. phases of flight. while the lateral dynamics are set to be faster than the original dynamics to satisfy performance specifications and to enforce MIL-standard requirements [72]. while their bandwidths are set to ensure a time-delay margin of at least 0. with a period of only a few seconds. Since the airplane is Level 1 at this flight condition. a linear prefilter is added to the adaptive flight control law to ensure desired decoupling 88 IEEE CONTROL SYSTEMS MAGAZINE properties as well as desired command-tracking performance. namely. the nominal desired dynamics of the linear state predictor are chosen to be similar to the linearized dynamics of the airplane. The L1 control law with its main elements is represented in Figure 25. These Mil-Spec guidelines are used in both civilian and military aircraft to shape dynamics associated with good handling qualities for both piloted and unmanned aircraft. and yaw rate for the lateral-directional dynamics. namely. The scale has ten points. Note that the rating is task specific. usually lightly damped.S.order equivalent systems. or one with full-authority command augmentation. For design purposes. Military Specifications (Mil-Spec) [72]. Flight Control Law Evaluation The L1 adaptive flight control law was evaluated on a twinturbine-powered. These specifications establish the relationships between pilot ratings. flight envelopes. thus neglecting velocity and pitch attitude. sideslip angle. the design of the L1 adaptive FCS for the GTM is based on the linearized dynamics of the aircraft at a nominal flight condition corresponding to an equivalent airspeed of 80 kt and an altitude of 1000 ft. Guidelines for stability and control of an aircraft. these linear dynamics are further simplified to include only shortperiod dynamics for the longitudinal axis. VA: AIAA. such as high speed above the allowable research envelope and post-stall high angle of attack. and a short-period oscillation. Aircraft Handling Qualities (AIAA Educational Series). a heavily damped aperiodic mode called the roll subsidence. with no scheduling or reconfiguration. where one indicates excellent and ten the worst qualities possible. Reston. which results in improved repeatability by leading the evaluation pilot through a series of decisions regarding the task performance and pilot workload. an uncontrollable aircraft resulting in a crash. Finally. and a short-period oscillation. are described in the U. usually with a period in seconds of about one-quarter of the airspeed in miles per hour. The evaluation of how well a piloted aircraft behavior follows these guidelines is based on the Cooper-Harper scale. The state predictor of the L1 control law is scheduled to specify different performance requirements at special flight regimes. although the aircraft may be controllable and flyable. which can be a divergence or subsidence. Hodgkinson. the adaptation sampling time is set to Ts 5 1 / 600 s. 1999. To improve the handling qualities of the airplane. known as low. the specific task rating might be poor. The lateral motion has three modes of motion. The stability characteristics of airplanes can be separated into longitudinal and lateral groups with the corresponding modes of motion. an aperiodic mode called the spiral mode. dynamically scaled GTM aircraft with tail number T2 during a series of flight tests conducted on » OCTOBER 2011 . which implies that.125 s and a gain margin of at least 6 dB. and adaptation rate are used across the entire flight envelope. called the Dutch-roll mode. We notice that the same control parameters for the prefilter. Overdamped second-order lowpass filters with unity dc gain are used in all control channels. The longitudinal modes of a statically stable airplane following a disturbance consist of a long-period oscillation called the phugoid oscillation. including the post-stall high angle-of-attack region. Handling qualities have a critical bearing on the safety of flight and on the ease of controlling an airplane in steady flight and in maneuvers. which corresponds to the execution speed of the AirSTAR flight control computer. aircraft class. The scale is dichotomous. and aircraft states. whether an open-loop airframe. these specifications provide specific guidelines on the desired dynamics for specific longitudinal and lateral motions of the aircraft as well as the H simplified dynamics that can express these dynamics. lowpass filters. The Cooper-Harper scale is universally used to enable the pilot to award a number to an aircraft to allow comparison with other aircraft or to show compliance with specifications. The latter is subjective but the task performance is reasonably repeatable provided that the task is well defined. Recall that L1 adaptive controller takes full advantage of the available computation speeds to improve the performance bounds. in the presence of faults and failures. Such a result would indicate a constraint on the aircraft performance and maneuverability and would typically be a result of a fault or failure. and roll rate. Some additional damping is added to both longitudinal and directional dynamics of the state predictor.Aircraft Handling Qualities and Their Specifications andling qualities involves the study and evaluation of the stability and control characteristics of an aircraft [S9].

for a given r0. define Lid ! d 1 d2 K 1 2. assume that. 0. x 1t 2. which is in the span of matrix Bm. z 1t 22 . ` for some r0 . there exist positive K1d. z 1t 22 (S50) 21 1 s 2 is a proper exponentially stable transfer matrix that C 1 s 2 Hm with dc gain C 1 0 2 5 Im. assume that there exist Lz. nonsingular. z 1t 222 1 Bum f2 1 t. X 2 ! fi 1 t. z 2 . and g : R 3 Rl 3 Rn S Rl are unknown nonlinear functions such that the standard assumptions on existence and uniqueness of solutions are satisfied. z 1 t 2 5 go 1 t. X2 2 7 ` # Kid 7 X12X2 7 `. that is. rin such that the L1-norm rr 2 7 Hxm 1 s 2 C 1 s 2 Kg 1s 2 7 L1 7 r 7 L`2rin L1rr rr 1 B0 (S53) condition 7 Gm 1s 2 7 L1 1 7 Gum 1s 2 7 L1. 7 x0 7 ` # r0 . i 5 1. whereas f2 [ represents the unmatched component. x 1t 2 . while f1 : R 3 Rn 3 Rp S Rm and f2 : R 3 Rn 3 Rp S R1n2m2 are unknown nonlinear functions that verify f1 1t. for all 7 Xj 7 ` # d. 0. for all v [ V. semi-strict-feedback systems. let g1 be an arbitrary small positive constant and. where xz 1t 2 [ Rl is the state vector of internal unmodeled dynamics. Hxum 1s 2 ! 1 sIn2Am 2 21Bum. 2. i 5 1. we assume that there exists a known compact convex set V such that v [ V ( Rm 3 m and that a nominal plant input gain v 0 [ V is known. x. 0 such that. Bz . which enters the plant dynamics orthogonally to the matrix Bm. Finally. z 1 t 22 . In this problem formulation. the class of systems considered includes general unmatched uncertainty that cannot be addressed by recursive design methods developed for strict-feedback systems. Also. 0 2 7 L` # Bi 0 for all t $ 0. xz 1 t 2 . Gum 1 s 2 ! 1 In2Hxm 1s 2 C 1s 2 Hm OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 89 . The signal z 1t 2 [ Rp is the output of the unmodeled dynamics # xz 1 t 2 5 g 1 t. The uncertain plant input gain matrix v is assumed to be an unknown. Bum 42 5 n. where Bi 0 . Bum 4 . In particular. 0 are known bounds. Bm [ Rn 3 m and C [ Rm 3 n are known full-rank constant matrices. f : R 3 Rn 3 Rp S Rn is an unknown nonlinear function.L1 Adaptive Controller for Systems with Unmatched Uncertainties he flight control system developed for the AirSTAR flight test vehicle uses an L1 adaptive controller designed for a class of multi-input. y 1 t 2 [ Rm is the regulated output. (S44) (S45) T Let X ! 3 x^. (S48) (S49) (S51) The design of the L1 adaptive controller involves a feedback gain matrix K [ Rm 3 m and an m 3 m strictly proper transfer matrix D 1 s 2 . 2. Hum 1 s 2 ! CHxum 1 s 2 5 C 1 sIn 2 Am 2 21Bum. (S46) (S47) Before we present the adaptive architecture and give an L1-norm stability condition. i 5 1. y 1 t 2 5 Cx 1t 2 . The choice of K and D 1 s 2 must ensure and. and also define rin ! 7 s 1 sI 2 Am 2 21 7 L1r0. 7 zt 7 L` # Lz 7 xt 7 L` 1 Bz. x 1t 2 . 2. f1 [ represents the matched component of the uncertainty. K2d such that 7 fi 1 t. there exists rr . for each d . z 1 t 22 c d 5 B21f 1t. f2 1t. z 1 t 22 . Hm 1 s 2 ! CHxm 1s 2 5 C 1 sIn2Am 2 21Bm. go : R 3 Rl S R p. without enforcing matching conditions [S10]. which lead. x 1 t 22 . Lz 1 d 1 g1 2 1 Bz 6 . v [ Rm 3 m is the uncertain plant input gain matrix. We rewrite (S44)–(S45) as # x 1t 2 5 Am x 1t 2 1 Bm 1 v u 1t 2 1 f1 1t. xz 1 0 2 5 xz 0. y 1 t 2 5 Cx 1 t 2 . for all t $ 0. Then let 7 fi 1 t. define Hxm 1 s 2 ! 1 s In 2 Am 2 21Bm. Consider the system with unmatched nonlinear uncertainty and with unmodeled dynamics given by # x 1 t 2 5 Am x 1t 2 1 Bm v u 1 t 2 1 f 1 t. where x 1 t 2 [ Rn is the state vector (measured). is satisfied. [S12]. The initial condition x0 is assumed to be inside a known set. Next. xz 1 t 22 . z^ 4 ^. and block-strict-feedback systems [S11]. strictly row-diagonally dominant matrix with sgn 1 v ii 2 known. j 5 1. we present an overview of this L1 control architecture. d 1 d 2 ! max 5 d 1 g1. Furthermore.0 . we assume that the transmission zeros of the transfer matrix Hm 1s 2 5 C 1 sI2Am 2 21Bm lie in the open left-half plane. x 1t 2 . and with a slight abuse of language let fi 1 t. x 1 t 2 . u 1 t 2 [ Rm is the control signal. Am [ Rn 3 n is a known Hurwitz matrix that defines the desired dynamics for the closed-loop system. x 1 0 2 5 x0. d id d Furthermore.and state-dependent unknown nonlinearities. 2. uniformly in t. B ! 3 Bm. to a strictly proper exponentially stable C 1s 2 ! v KD 1 s 2 1 Im 1 v KD 1 s 22 21 (S52) x 102 5 x 0. 0. where r0 5 max 7 x0 7 `. where Bum [ Rn 3 1n2m2 is a constant matrix such that B^ m Bum 5 0 and also rank 1 3Bm. Next. for each d . X1 2 2fi 1 t. where Gm 1s 2 ! Hxm 1s 2 1 Im2C 1 s 22 . 21 1 s 2 C 2 Hxum 1 s 2 . x 1t 2. pure-feedback systems. multi-output uncertain plants in the presence of uncertain system input gain and time.

[S12] B. B is introduced time. Nonlinear and Adaptive Control Design. if 1 21 we choose Kg 1 s 2 as the constant matrix Kg 5 2 1 CA2 . then the diagonal entries of the desired transfer matrix M 1s 2 5 C 1sIn2Am 2 21Bm Kg have dc gain equal to one. z 1 t 22 . x ^ 1 02 5 x0. Tomizuka. and with h ^ 1 1t 2 ! s ^ 1 1t 2 estimating the matched h uncertainty. u ref 1s 2 5 kD 1 s 2 Hm ^ 1 1 t 2 [ Rm and s ^ 2 1 t 2 [ Rn2m are the adaptive estimates where s of the matched and unmatched uncertainties. the tracking error between y 1 t 2 and yref 1t 2 . As an example. Baltimore. xref 1 0 2 5 x0. both in transient and steady state. is uniformly bounded by a constant that can be arbitrarily reduced by decreasing the adaptation sampling time Ts. 1 i11 2 Ts 2 . B0 ! max e B10. “Adaptive robust control of SISO nonlinear systems in a semi-strict feedback form. uref 1s 2 5 2v 21C 1s 2 1 h1ref 1s 2 1 Hm (S60) 90 IEEE CONTROL SYSTEMS MAGAZINE » OCTOBER 2011 . THEOREM 1] Let the feedback gain K and the filter D 1 s 2 be chosen to satisfy the L1-norm condition in (S53).1.while . 1995. Xargay.” where we achieve arbitrary performance bounds by choosing an appropriate value of the sampling time Ts. 5. y (S55) where hi ref 1s 2 is the Laplace transform of hi ref 1t 2 ! fi 1t. “L1 adaptive controller for multi-input multi-output systems in the presence of nonlinear unmatched uncertainties.0 f. vol. where Ts . g1 1 Ts 2 . Therefore. c. pp. Cao. Again. THEOREM 1 [S10. no. m 1e while m 1 Ts 2 5 e AmTs. Krsti´ c. Kokotovi´ c. z 1 t 22 . The L1 reference system for this architecture is given by # xref 1t 2 5 Amxref 1t 2 1 Bm 1 v uref 1 t 2 1 f1 1 t. Due to this structural equivalence between the L1 reference controller and a linear internal model controller. arbitrarily close tracking performance can be achieved. Yao and M. V. These results are consistent with the results in “L1 Adaptive Output-Feedback Controller with Piecewise Constant Adaptation Laws. MD. 2.0 ! L2rr L1rr . 21 ^ 2m 1s 2 ! Hm ^ 2 1 s 2 captur1s 2 Hum 1s 2 h with rg 1 s 2 ! Kg 1s 2 r 1 s 2 . The state-predictor is given by # ^ 1t 2 5 Am x ^ 1t 2 1 Bm 1 v0 u 1t 2 1 s ^ 1 1 t 22 1Bum s ^ 2 1t 2 . robust control methods can be applied to design a lowpass filter that satisfies the L1 norm stability condition in (S53). there exists an adaptation sampling time Ts such that |7 # g 1T 2. x in (S50). and P. New York: Wiley. and 1 AmTs F 1Ts 2 ! A2 2I 2 . 1997. Next. g1 1 Ts 2 . h (S58) (S57) 7 uref 2 u 7 L` # g2 1 Ts 2 . The control law is given by ^ 1s2. Notice that conventional design methods from multivariable control theory can be used to design the prefilter Kg 1 s 2 to achieve desired decoupling properties. The next theorem summarizes the main result for this adaptive architecture. B20 . i 5 1. m Bm 2 off-diagonal entries have zero dc gain. d 5 2c d B 21F21 1 Ts 2 m 1Ts 2 x s s 0 In2m ^ 21t2 t [ 3 iTs . 0 is the adaptation sampling | 1t 2 5 x ^ 1t 2 2x 1t 2 is the prediction error. 33. yref 1 t 2 5 Cxref 1t 2 . using the ideas discussed in “Bridging Adaptive and Robust Control.” we can derive the reference controller from the equations of the L1 reference system (S59)–(S60) to arrive at 21 1 s 2 1 Hm 1s 2 Kg 1 s 2 r 1 s 2 2 Cxref 1 s 22 . bounded-output stable. (S61) and Kg 1 s 2 is the exponentially stable feedforward prefilter. where ^ 2 1t 2 ! s ^ 2 1t 2. and g2 1 Ts 2 are known class K functions of Ts. and g2 1 Ts 2 can be (S59) 21 1 s 2 Hum 1 s 2 h2ref 1s 2 2Kg 1s 2 r 1 s 22 . For details see [S10]. u 1s 2 5 2KD 1s 2 h ^ 1s 2 is the Laplace transform of the signal where h ^ 1 t 2 ! v 0u 1t 2 1 h ^ 1 1t 2 1 h ^ 2m 1 t 2 2rg 1t 2 . the reference system in (S59)–(S61) is bounded-input.. and C.” Automatica. as well as u 1 t 2 and uref 1 t 2 . 893–900. REFERENCES [S10] E. Thus. we introduce the elements of the L1 adaptive control architecture. z 1 t 222 1 Bumf2 1t. where g0 1Ts 2 . 874–879. 2010. I. Then. Kanellakopoulos. (S54) x ^ 1 t 2 5 Cx ^ 1t 2. Moreover. Hovakimyan. (S56) which is similar to the reference controller for the state feedback with matched uncertainty in (S7). 2. while the constructed based on knowledge of the uncertainty bounds and the adaptive control system parameters. for both signals simultaneously by adapting sufficiently fast. Notice that the functions g0 1Ts 2 . pp. 7x L` 0 s (S62) (S63) (S64) for i 5 0. xref 1 t 2 . N. 7 xref 2 x 7 L` # g1 1 Ts 2 . xref 1 t 2 . respectively. These estimates are driven by the adaptation laws c s 0 Im ^ 1 1t 2 | 1 iT 2 . h ing the estimation of the unmatched uncertainty.” in Proc. American Control Conf. xref 1t 2 . [S11] M.

and handling qualities assessment based on Cooper-Harper ratings (CHR) as well as subjective pilot comments. The injection of the wavetrain is indicated in the plots with a purple dashed line. both the longitudinal and lateral stability of the aircraft are simultaneously degraded. the control law uses lowpass filters to generate control-surface deflection commands d.5 0 –0.and roll-doublet wavetrains. as well as the ability of the adaptive controller to maintain aircraft stability and predictable performance in off-nominal adverse flight conditions as a measure of robust performance.5 –1 –1.June 4. The time-delay margin is determined by progressively increasing the latency in the control channel until instability occurs.5 Sideslip Angle (deg) 1 0. 2010 at Ft. and part of another turn. To place the flight test results into the appropriate context. The L1 control law consists of an estimation scheme and a control law. the pilot is asked to hold the aircraft in controlled flight at an αcmd pcmd βcmd L1 Flight Control Law Pilot Command Prefilter Control Law δ GTM x Aircraft x State − Predictor ˜ x η Adaptation Law ∧ ∧ FIGURE 25 Block diagram of the L1 flight control architecture. In the stability degradation tasks. 15 Pitch Axis (deg. (c) slideslip-angle tracking. The adaptation law signal x updates the estimates of the uncertainties present in the plant at a high adaptation rate. These plots show 70 s of flight including angle-of-attack and roll-rate wavetrains. OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 91 . The flight control law evaluation also includes high angle-of-attack flight and recovery in the stall and post-stall region of the flight envelope. are used to excite the closed-loop system dynamics and evaluate tracking performance. The estimation scheme includes a state predictor and an adaptation law. deg/s) 50 Roll Axis (deg. (b) roll-axis response. a summary of the flight test procedures is presented. Based on the estimates of the uncertainties. and (d) control surface commands. which are superimposed on the pilot’s stick commands. post-stall flight and recovery. Pickett. In these tasks. Virginia. 2010 and September 10. In particular. The state predictor generates a prediction x present in the plant h of the state x. This nominal response is considered to have solid Level 1 handling qualities according to pilot ratings and comments. while reducing pitch-control authority by 50%. The performance of the adaptive controller is evaluated through a series of tasks designed to test the LOC prevention and recovery capabilities of the closed-loop adaptive system. deg/s) 10 5 0 –5 α cmd α q 0 pcmd p φ –10 420 430 440 450 460 470 480 490 Time (s) (a) β cmd β Surface Commands (deg) –50 420 430 440 450 460 470 480 490 Time (s) (b) 10 5 0 –5 Rudcmd Elecmd Ailcmd –10 420 430 440 450 460 470 480 490 Time (s) (d) 1. (a) Pitch-axis response. The difference between these signals yields an error | that drives the adaptation process.5 420 430 440 450 460 470 480 490 Time (s) (c) FIGURE 26 Pitch and roll doublets for the nominal aircraft. which are used to generate estimates of the uncertainties ^ ^ . The main objective of the flight control law evaluation is to assess the time-delay margin of the adaptive system as a measure of closed-loop robustness. one complete turn. The pitch. these tasks include flight under aerodynamic stability degradation.

the stability is reduced in decrements of 25% until the closed-loop system Robustness Evaluation To determine the time-delay margin of the adaptive closedloop system. with a nominal system latency of approximately 0. deg/s) q Elevator Deflections (deg) α cmd α 6 4 2 0 –2 –4 425 10 5 0 –5 –10 425 EleIL EleOL EleIR EleOR Elecmd 430 Time (s) (a) 435 440 430 Time (s) (b) 435 440 4 Roll and Yaw Rates (deg/s) Aileron Deflections (deg) 10 5 0 –5 –10 425 pcmd p r 3 2 1 0 –1 –2 –3 –4 425 Ailcmd 430 AilR 435 Time (s) (d) 440 430 Time (s) (c) 435 440 FIGURE 27 Pitch-doublet response for the nominal aircraft. which provides a first-order-like angle-of-attack response. pilot assessment is also provided for off-nominal cases with simultaneous stability degradation in both longitudinal and lateral axes. starting with a 50% degradation. This particular task is considered to be demanding for a pilot under nominal conditions. References [7] and [8] provide additional data and analysis of the results. the roll and yaw rates remain within the turbulence level. (a) Pitch-axis response. a brief overview of representative results from the flight tests conducted in June and September 2010 is presented. For this task. during the injection of the wavetrain. which. Finally. the pilot is asked to fly the aircraft in a figure-eight pattern at approximately 900 ft and 80 kt. angle of attack above 13° while keeping the bank angle below 45°. In fact.022 s. handling qualities assessment involves highworkload offset landing tasks. The initial task condition is a final approach at 80 kt characterized by an offset from the center line of the virtual runway that is 100 ft to the left of the centerline. the control channel until the aircraft response becomes unstable or presents sustained high-amplitude oscillations.147 s. 100 ft above the runway.15 Pitch Axis (deg.125-s latency added to the control channel. amounts to a total time-delay margin of approximately 0. ensuring reduced cross-coupling between channels. During the flight test evaluation conducted in June 2010. The injection of the wavetrain is indicated in the plots with a purple dashed line. The flight control law. while pitch and roll doublets are superimposed on the pilot’s stick commands to excite the system dynamics. Flight Under Aerodynamic Stability Degradation The evaluation conducted in June 2010 was also used to determine the ability of the L1 flight control law to compensate for simultaneous longitudinal and lateral aerodynamic stability degradation and a 50% reduction in pitch-control authority. (b) elevator deflections. (c) lateral-directional response. In the following section. and (d) right-aileron deflection. To evaluate the performance of an adaptive controller. is able to isolate the commanded response to the longitudinal axis. A description of this task with the corresponding flight data can be found in [7]. a progressively increasing latency is added to 92 IEEE CONTROL SYSTEMS MAGAZINE » OCTOBER 2011 . and 1800 ft downrange from a desired touchdown box marked on the runway. it was determined that the L1 flight control law is robust to a 0. where the pilot is asked to land the aircraft in a virtual scenario presented on the pilot’s head-up display. The stability degradation task is introduced gradually.

the pilot is asked to fly the aircraft in a figure-eight pattern at approximately 900 ft and 80 kt. deg/s) β r 1 0. (e) pitch-axis response. which correspond. Figure 26 shows 70 s of flight data including the pitch and roll wavetrains. which is considered to have solid Level 1 handling qualities according to pilot ratings and comments. while pitch. (c) directional response. The injection of the wavetrain is indicated in the plots with a purple dashed line. while maintaining the angle of attack and sideslip angle within 60.5 465 Rudcmd Rud 5 0 –5 –10 465 470 Time (s) (c) 475 480 470 Time (s) (d) 475 480 15 6 Elevator Deflections (deg) Pitch Axis (deg. deg/s) p φ 6 4 2 0 –2 –4 465 Ailcmd AilR 0 –50 465 470 Time (s) (a) 475 480 470 Time (s) (b) 475 480 10 Rudder Deflections (deg) β cmd 1. one complete turn. to a neutrally stable aircraft and an unstable aircraft.5° of the commanded values.5 –1 –1. the nominal response of the aircraft with the L1 flight control law. Of particular interest are the cases of 100% and 125% stability degradation.5 Sideslip Angle and Yaw Rate (deg. and part of another turn.5 0 –0.and roll-doublet wavetrains are superimposed on the pilot’s stick commands during the straight legs of the pattern to excite the system dynamics. (a) Roll-axis response. with the tested flight controller is unstable. respectively.5 –2 –2. (d) rudder deflection. (b) right-aileron deflection. Similar to the latency-injection task.Aileron Deflections (deg) 50 8 pcmd Roll Axis (deg. The L1 controller tested in June 2010 was able to maintain aircraft stability for up to 100% simultaneous degradation in longitudinal and lateral aerodynamic stability with an additional 50% reduction in pitch-control authority. is presented. The flight control law provides a predictable first-order-like roll-rate response. To provide a basis for comparison. and (f) elevator deflections. deg/s) α cmd 10 α q 4 2 0 –2 –4 465 EleIL EleOL EleIR EleOR Elecmd 5 0 465 470 Time (s) (e) 475 480 470 Time (s) (f) 475 480 FIGURE 28 Roll-doublet response for the nominal aircraft. The wavetrains can be observed in Figure 26(a) and (b) as precise straight-line OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 93 .

which is considered a severe failure. there is more aileron activity to compensate for turbulence-induced roll rate. see Figure 27(c). see Figure 28(c) and (e). Figure 27 shows the pitch-axis wavetrain with the corresponding lateral-directional response. The main difference is a higher control surface activity to compensate for turbulence. during the injection of the roll-axis wavetrain. reported that he did not detect a perceptible difference from flying the nominal aircraft. while Figure 28 presents the roll-axis wavetrain with the corresponding longitudinal response. is able to isolate the commanded response to the appropriate axis. the closed-loop system » OCTOBER 2011 . deg/s) q 0 α cmd 590 600 610 620 α 630 640 –50 580 pcmd 590 p 600 610 φ 620 630 640 Time (s) (a) 50 15 Time (s) (b) Surface Commands (deg) 10 5 0 –5 Rates (deg/s) 0 –50 580 p 590 q 600 610 r 620 630 640 Elecmd –10 580 590 Ailcmd 620 Rudcmd 630 640 600 610 Time (s) (c) Time (s) (d) FIGURE 29 Pitch and roll doublets for the case of 75% stability degradation.5° of the commanded values. The control-surface deflection commands for this flight segment are illustrated in Figure 26(d). During the injection of the pitch-axis wavetrain. the roll and yaw rates do not exceed 65 deg/s. which provides a first-order-like response for both channels. however. The injection of the wavetrain is indicated in the plots with a purple dashed line. The reduction in longitudinal stability is obtained by closing a destabilizing feedback loop from angle of attack to inboard elevators. (c) body rates. and (d) control-surface commands. both the angle of attack and sideslip angle are maintained within 60. The flight control law. Recall that the stability degradation is applied simultaneously in the longitudinal and lateral axes while the elevator effectiveness is reduced by 50%. In addition. ensuring reduced cross-coupling between channels. deg/s) 10 5 0 –5 –10 580 Roll Axis (deg. The results for the case of 50% stability degradation are similar to the results for the nominal aircraft shown in figures 26–28. the performance of the closed-loop adaptive system in the presence of aerodynamic stability degradation is analyzed. The main comment from the 94 IEEE CONTROL SYSTEMS MAGAZINE pilot during the post-flight debrief is that the L1 flight control law provides “a nice Level 1 flying airplane for light turbulence conditions. while the injection of the stability degradation is indicated with a cyan dashed line. which is considered to be within the turbulence level. As shown in the figures. doublets. The pilot. which indicates that the wavetrain injection is done with the pilot’s hands off the stick. With respect to the case of nominal aircraft. while the spoilers are used asymmetrically to reduce lateral stability through roll damping. is illustrated in figures 29–31. The response of the closed-loop system to the pitch and roll wavetrains can be observed in detail in figures 27 and 28. while the pitch-axis response remains similar. Figure 26(c) shows that the L1 adaptive-control law maintains the sideslip-angle tracking error within 60. The case of 75% stability degradation. (a) Pitch-axis response. the closed-loop system experiences a slight degradation in roll-rate tracking performance. In particular.” Next. for example. The results for the 50% degradation case can be found in [7].8°. especially in the roll channel.15 50 Pitch Axis (deg. (b) roll-axis response. the constant nonzero sideslipangle command is due to the vane-calibration bias. Similarly.

since they are used to produce the 75% longitudinal stability degradation. The left and right inboard elevators do not track the elevator command provided by the L1 control law.15 Pitch Axis (deg. the L1 flight control law is not able to maintain prescribed performance. Figure 32 shows angle-of-attack and roll-rate tracking as well as the angular-rate response and the corresponding actuator commands through the entire segment of flight. and (d) right-aileron deflection. (a) Pitch-axis response. Flight test results for this case can be found in [7]. Moreover. the inboard elevators are scheduled with angle of attack to create a destabilizing feedback loop and thus do not follow the elevator command generated by the L1 flight control law. as shown in Figure 33(c). (b) elevator deflections. For the case of 100% stability degradation. as evidenced by overshoot in the angle-of-attack response and larger amplitudes in pitch rate. there is an interest OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 95 . interest in stall and post-stall flight is twofold. Stall and Post-Stall Flight From the perspective of the IRAC project. there is more aileron activity to compensate for the turbulenceinduced roll rate. resulting in an evident deterioration in the closed-loop system performance. The response of the aircraft to the pitch and roll wavetrains are isolated in figures 33 and 34. The flight control law provides a first-order-like angle-ofattack response similar to the response of the nominal aircraft. The injection of the wavetrain is indicated in the plots with a purple dashed line. First. (c) lateral-directional response. The roll response exhibits oscillatory behavior as well as increased aileron activity. deg/s) q Elevator Deflections (deg) α cmd 6 α 10 5 0 –5 –10 580 4 2 0 –2 –4 580 EleIL EleOL EleIR EleOR Elecmd 585 590 595 Time (s) (a) 600 605 585 590 595 Time (s) (b) 600 605 4 Roll and Yaw Rates (deg/s) Aileron Deflections (deg) 10 5 0 –5 –10 580 pcmd p r 3 2 1 0 –1 –2 –3 –4 580 Ailcmd 585 AilR 590 595 Time (s) (d) 600 605 585 590 595 Time (s) (c) 600 605 FIGURE 30 Pitch-doublet response for the case of 75% stability degradation. experiences a slight degradation in roll-rate tracking performance. Also. there is more aileron activity to compensate for turbulence-induced roll rate. respectively. Figure 33 shows that the pitch-axis closed-loop damping is reduced. In Figure 33(b). the angle-of-attack and sideslip-angle tracking errors are maintained within 61 deg/s. despite the controller performance degradation in the roll axis. in fact. while the injection of the stability degradation is indicated with a cyan dashed line. In addition. In particular. Due to the reduction in roll damping. the pitch-axis doublet induces 615-deg/s oscillations in roll rate. During the injection of the roll-axis wavetrain. The L1 flight control law flown during the June 2010 deployment was not able to handle the case of 125% stability degradation. which seems to indicate that the controller is reaching the limit of its ability to provide robust performance in the roll channel. The pilot observed roll ratcheting during the injection of the pitch-axis wavetrain. while the pitch-axis response remains similar to the response of the nominal aircraft. see Figure 34. as mentioned above. the use of the inboard elevators to produce the 100% longitudinal stability degradation is evident. the closed-loop system exhibits large overshoot and 620-deg/s sustained oscillations in roll rate.

Second. roll damping characteristics are expected to rapidly change from stable to unstable in the 10–12 ° angle-of-attack range.5 Rudder Deflections (deg) Sideslip Angle and Yaw Rate (deg. and (f) elevator deflections. at post-stall 15 ° angle of attack. roll damping is expected to improve and aerodynamic asymmetry is expected to decrease. Roll 96 IEEE CONTROL SYSTEMS MAGAZINE asymmetry with an abrupt wing drop is expected to occur at an angle of attack of about 13°. moreover. The roll-rate response exhibits an overshoot. deg/s) φ 8 6 4 2 0 –2 –4 –6 0 –50 625 Ailcmd 630 635 Time (s) (b) 640 AilR 645 630 635 Time (s) (a) 640 645 625 10 1. high angle-of-attack flight provides a challenging scenario for adaptive controllers in which the dynamics of the aircraft experience significant changes and are characterized by uncertain nonlinear phenomena. while the injection of the stability degradation is indicated with a cyan dashed line. The primary challenge associated with post-stall high angle of attack is rapidly changing roll damping characteristics and an abrupt wing drop due to asymmetric wing stall. (a) Roll-axis response. (c) directional response. (d) rudder deflection. The injection of the wavetrain is indicated in the plots with a purple dashed line. (b) right-aileron deflection.Aileron Deflections (deg) 50 pcmd p Roll Axis (deg. (e) pitch-axis response. These plots illustrate the degradation in roll-rate tracking performance caused by the reduction in roll damping. The defining stall characteristic is expected to be an unstable pitch break occurring at an angle of attack of 13°. deg/s) α cmd 10 α q 4 2 0 –2 –4 625 EleIL EleOL EleIR EleOR Elecmd 5 0 625 630 635 Time (s) (e) 640 645 630 635 Time (s) (f) 640 645 FIGURE 31 Roll-doublet response for the case of 75% stability degradation. deg/s) β cmd 5 0 β r 1 0.5 –2 –2. which reduces the predictability of the closed-loop response. in exploiting the AirSTAR facility as a flying wind tunnel to model unsteady nonlinear aerodynamics encountered at post-stall flight regimes.5 –1 –1.5 0 –0. static and dynamic longitudinal stability is expected to be low. » OCTOBER 2011 .5 625 630 Rudcmd Rud –5 –10 625 630 635 Time (s) (c) 640 645 635 Time (s) (d) 640 645 15 6 Elevator Deflections (deg) Pitch Axis (deg. Also. At angles of attack above stall.

maintaining the roll rate within 620 deg/s and the bank angle within 615 °. post-stall regime on the GTM aircraft. deg/s) 0 pcmd p φ 680 690 700 710 Time (s) (b) 720 730 α cmd 680 690 α q 720 730 –50 700 710 Time (s) (a) 50 Surface Commands (deg) 10 5 0 –5 –10 Rates (deg/s) 0 p –50 680 q 690 r 700 710 Time (s) (c) 720 730 Elecmd 680 690 Ailcmd 700 710 Time (s) (d) Rudcmd 720 730 FIGURE 32 Pitch and roll doublets for the case of 100% stability degradation. The pilot captures and maintains the angle of attack at 18° for more than 7 s. Figure 35 shows the result of three unsuccessful attempts to capture an angle of attack of 18° without exceeding 45° of bank angle. The roll response exhibits oscillatory behavior as well as increased aileron activity. the pilot had the confidence to take the GTM into post-stall flight during the June 2010 deployment. Figure 36 shows the results of two attempts to capture an angle of attack of 18° without exceeding the prescribed limit of 45° of bank angle. The time-history response for one of these attempts is illustrated in Figure 36(c). while the bank angle exceeds the 45 ° limit. the combination of power setting and pilot command saturates all elevator segments from 849 to 854 s without observable ill effects on the FCS or aircraft dynamics. recovers by 856 s and initiates a turn by 858 s. see Figure 36(d). the pilot initiates post-stall recovery. To date. Based on the performance and predictability of the L1 flight control law observed in the stability degradation tasks. there is an abrupt roll-rate response exceeding 40 deg/s. while the injection of the stability degradation is indicated with a cyan dashed line. the bank-angle excursion lasts for less than 2 s before returning to a 5 ° offset. Moreover. 15 50 Pitch Axis (deg. During these attempts. At 854 s.Fast and robust adaptation is the key feature of the L1 adaptive control that leads to a priori quantifiable transient and steady-state performance. the L1 flight control law is the only control law that has flown in the high angle-of-attack. Figure 35 presents flight test results for a post-stall angle-of-attack acquisition task. and (d) control-surface commands. (c) body rates. In particular. (b) roll-axis response. OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 97 . The injection of the wavetrain is indicated in the plots with a purple dashed line. which illustrates some of the stall phenomena described above. In all three attempts. (a) Pitch-axis response. deg/s) 10 5 0 –5 –10 Roll Axis (deg. the L1 flight control law exhibits a predictable and repeatable behavior. resulting in an evident deterioration in closed-loop system performance. For this particular attempt.

(c) lateral-directional response. this rating deteriorates to CHR 4. Moreover. the implication of turbulence is that the control surface deflections required to make aircraft neutrally stable. the pilot achieves a safe landing with CHR 7. Furthermore. Under nominal aircraft stability. the task is performed not only for the nominal aircraft but also for the cases of 100% and 125% stability degradation with an additional 50% reduction in pitch-control authority. Level 2. For the case of 100% stability degradation. flight-path angle. The L1 flight control law provided a graceful degradation in terms of flying qualities rating for each of the stability degradation cases. We refer to [8] for the flight test data of these offset landing tasks. nominal CHR 3. Offset Landings The L1 flight control law was evaluated in a series of offset landing tasks. Level 1 handling qualities. 100% stability degradation. 100% roll. in reality. Level 1 flying qualities. Level 3 handling qualities. the task requirements are set on aircraft attitude. which were conducted during the September 2010 deployment. To place the results in an appropriate context. as mentioned above. (a) Pitch-axis response. to evaluate the ability of the adaptive controller. note that the GTM. The injection of the wavetrain is indicated in the plots with a purple dashed line. Hence. which typically exacerbates stability degradation especially in roll axis since the destabilizing feedback loop responds to turbulence induced roll rate. deg/s) q Elevator Deflections (deg) α cmd 10 α 8 6 4 2 0 –2 –4 –6 10 5 0 –5 –10 EleIL EleOL EleIR EleOR Elecmd 675 680 685 Time (s) (a) 690 675 680 Time (s) (b) 685 690 20 Roll and Yaw Rates (deg/s) 15 10 5 0 –5 –10 –15 –20 Aileron Deflections (deg) pcmd p r 6 4 2 0 –2 –4 –6 675 680 Time (s) (d) 685 690 Ailcmd AilR 675 680 685 Time (s) (c) 690 FIGURE 33 Pitch-doublet response for the case of 100% stability degradation. be larger and make the aircraft slightly unstable. is rated CHR 3. and touchdown position on the runway. and unstable CHR 7. for the neutral stability case. The injection of the pitch-axis wavetrain induces 615-deg / s oscillations in roll rate. and (d) right-aileron deflection. aircraft uncontrollable with catastrophic landing expected. These plots show that the pitch-axis closed-loop damping is reduced. as evidenced by an overshoot in the angle-of-attack response and larger amplitudes in pitch rate. A summary of comments and CHRs that the AirSTAR research pilot » OCTOBER 2011 .15 Pitch Axis (deg.and pitch-axis stability degradation. Finally. without any feedback FCS. The offset landing task is considered to be a high workload task in which the pilot tries to put the aircraft down in a desired touchdown spot on the runway. while the injection of the stability degradation is indicated with a cyan dashed line. would. that is. the L1 control law is rated CHR 3. (b) also illustrates the use of the left and right inboard elevators to produce the 100% longitudinal stability degradation. the L1 controller experiences some handling qualities degradation and is rated CHR 5. neutrally stable CHR 5. 98 IEEE CONTROL SYSTEMS MAGAZINE The offset landing task was flown under light turbulence conditions. (b) elevator deflections. the pilot rated the task as CHR 10. for the case of 125% stability degradation. Level 2 flying qualities for the offset landing task of a nominal aircraft. During the flight test. that is.

L1 Flight Control Law in Support of Unsteady Aerodynamic Modeling Work By the September 2010 deployment.5 0 –0. The injection of the wavetrain is indicated in the plots with a purple dashed line. the L1 adaptive flight control law had established itself as a reliable and predictable tool to be used in support of other research tasks to reduce the pilot’s workload and provide tighter acquisition of target flight conditions. while the injection of the stability degradation is indicated with a cyan dashed line. Moreover.5 –1 –1.5 1 0. the control law demonstrates a precision tracking capability at the edge of aircraft controllability even with persistently saturated control surfaces. Significantly. (e) pitch-axis response. which seems to indicate that the controller is reaching the limit of its ability to provide robust performance in the roll channel. deg/s) β cmd 5 0 β r 1. During the injection of the roll-axis wavetrain.Aileron Deflections (deg) 50 pcmd p Roll Axis (deg.5 –2 –2. One of the research tasks flown during the September 2010 deployment was the calibration of the two air-data vanes placed on each wingtip of the GTM aircraft. (a) Roll-axis response.5 710 715 Rudcmd Rud –5 –10 710 715 720 Time (s) (c) 725 730 720 Time (s) (d) 725 730 15 8 Elevator Deflections (deg) Pitch Axis (deg. (b) right-aileron deflection. (c) directional response. (d) rudder deflection. and (f) elevator deflections. provided for the offset landing task evaluation can be found in [78]. The precise calibration of the angle-of-attack and angle-of-sideslip vanes is critical for system identification and nonlinear aerodynamic modeling. the L1 control law provides predictable response to the pilot when the control authority for different axes is divided between a OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 99 . deg/s) φ 8 6 4 2 0 –2 –4 –6 –8 710 715 0 –50 710 Ailcmd 720 Time (s) (b) 725 AilR 730 715 720 Time (s) (a) 725 730 10 Rudder Deflections (deg) Sideslip Angle and Yaw Rate (deg. deg/s) 10 5 0 –5 6 4 2 0 –2 –4 710 EleIL EleOL EleIR EleOR Elecmd α cmd –10 710 715 α 720 Time (s) (e) q 725 730 715 720 Time (s) (f) 725 730 FIGURE 34 Roll-doublet response for the case of 100% stability degradation. the system response exhibits overshoots and sustained oscillations in roll rate.

Angle of Attack (deg) 20 Acquisition 1 Acquisition 2 Acquisition 3 40 20 0 –20 –40 –60 590 α (deg) 15 10 5 0 p 595 φ α 600 Time (s) (d) 605 610 –6 –4 –2 0 2 β (deg) (c) 4 6 FIGURE 35 Stick-to-surface post-stall angle-of-attack acquisition task. For this modeling work. These plots show the result of three unsuccessful attempts to capture an 18° angle of attack without exceeding 45° of bank angle. which provided time for the optimized multi-input wavetrains to excite the aircraft dynamics in all 6DOF to collect the data needed for real-time dynamic modeling. the tradeoff for increased performance led to a decrease in closed-loop latency robustness. and (d) time-history response (single attempt). (b) roll rate.147 s mentioned above to 0. The L1 control operated without introducing high correlations in the states and controls during the maneuvers. Figure 37(a) illustrates a precision tracking task in which the GTM is required to track a multistep wavetrain in the post-stall regime. the roll rate experiences an abrupt departure and exceeds 40 deg/s. The L1 control law allowed the research pilot to operate the aircraft in precarious flight conditions near stall and departure for longer periods of time. during the May 2011 deployment. deg/s). The results of these flight tests are presented in [76].117 s. (a) Bank angle. also in the post-stall regime. the L1 flight control law provides high angle100 IEEE CONTROL SYSTEMS MAGAZINE of-attack tracking capabilities and ensures a safe recovery once the post-stall maneuver is completed. In all three attempts. which occurs at about 13° angle of attack. while the bank angle exceeds the 45° limit. track a predetermined input wavetrain with the pilot’s hands off the stick. As shown in Figure 37(b). Flight test results for two of these modeling tasks are shown in Figure 37. In addition. which decreased from the 0. To improve tracking performance at high angles of attack to better support these modeling tasks. Further description of the performance of the L1 adaptive flight control law on these and other supporting tasks can be found in [8].25 20 Acquisition 1 Acquisition 2 Acquisition 3 25 20 Acquisition 1 Acquisition 2 Acquisition 3 α (deg) 15 10 5 0 α (deg) –60 –40 –20 0 20 φ (deg) (a) 40 60 15 10 5 0 –50 0 p (deg/s) (b) 50 25 60 Roll Axis (deg. In particular. the L1 adaptive flight control law was » OCTOBER 2011 . it was necessary to increase the bandwidth of the L1 lowpass filter in the longitudinal channel. (c) sideslip angle. pilot and an automatic command. Moreover. the L1 adaptive-control law continued to enable exploration of departure-prone edges of the flight envelope and support nonlinear unsteady aerodynamic modeling beyond the linear flight regime. As predicted by L1 adaptive-control theory. The L1 adaptive flight control law was used as a means to maintain departure resistance and to reduce pilot workload during test-point acquisition. the L1 flight control law was used to support unsteady aerodynamic modeling work at post-stall. just upstream of the rate and position limiting. Note that the maneuvers were implemented by adding wavetrain perturbation inputs directly to the actuator commands from the control law. Finally. The primary challenge associated with post-stall high angle of attack is rapidly changing roll damping characteristics and an abrupt wing drop due to asymmetric wing stall. once the desired angle of attack is reached. during the September 2010 deployment. while Figure 37(b) presents the response of the system to a Schroeder sweep. the main objective of the control law is to stabilize the aircraft at an angle of attack of 18° and.

deg/s) Angle of Attack (deg) 30 20 10 0 –10 –20 –30 830 p φ α 0 –5 –10 –15 –20 –25 –30 –35 830 Elecmd EleIL EleOL EleIR EleOR 835 840 845 850 Time (s) (d) 855 860 835 840 845 850 Time (s) (c) 855 860 FIGURE 36 Post-stall angle-of-attack acquisition task with the L1 adaptive flight control law. in which the aircraft is required to track. The combination of power setting and pilot command saturates all of the elevator segments from 849 to 854 s without observable ill effects on the control system or aircraft dynamics. In these modeling tasks. OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 101 . (a) a multistep wavetrain and (b) a Schroeder sweep. (c) time-history response (single attempt). the L1 flight control law exhibits a predictable repeatable behavior. (b) roll rate. 22 22 Angle of Attack (deg) Angle of Attack (deg) 20 18 16 14 12 10 8 6 4 650 αcmd α 20 18 16 14 12 10 8 6 4 740 αcmd α 655 660 665 670 Time (s) (a) 675 680 750 760 Time (s) (b) 770 780 FIGURE 37 L1 flight control law in support of unsteady aerodynamic modeling work in post-stall regimes. and (d) elevator deflections. maintaining the roll rate within 620 deg / s and the bank angle within 615°. once the desired angle of attack is reached. Flight tests results for these research tasks are reported in [8]. The purpose of this maneuver was to validate unsteady aerodynamic behavior. respectively.25 20 Acquisition 1 Acquisition 2 25 20 α (deg) 15 10 5 0 α (deg) 15 10 5 0 Acquisition 1 Acquisition 2 –60 –40 –20 0 20 φ (deg) (a) 40 60 –50 0 p (deg/s) (b) 50 40 5 Elevator Deflections (deg) Roll Axis (deg. track a predetermined input wavetrain with the pilot’s hands off the stick. (a) Bank angle. The control law was used to track roll-axis excitation waveforms in flight conditions near and post stall. the main objective of the control law is to track an angle of attack of 18° and. During these attempts. used to support an unsteady aerodynamics modeling investigation at aerodynamic stall. which was identified in wind-tunnel tests using a forced-oscillation test technique. These plots illustrate two precision tracking tasks.

control in the presence of limited information. The results show that the L1 flight control law provides predictable behavior both in the neighborhood of the design point as well as at other operating points of the flight envelope. 1206 West Green Street. She received a B. France.S. in aerospace engineering from the Politecnico di Torino. by facilitating precision tracking with reduced pilot workload. networks of autonomous systems and game theory. majoring in optimal control and differential games.D. in manufacturing engineering from Boston University in 1995 and 1999. in physics and mathematics in 1992. and mechatronics.S. and cooperative control of autonomous systems. Her research includes advanced » OCTOBER 2011 . IL 61801 USA. and verifies the theoretical predictions of its behavior. he was a research scientist in the Department of Aerospace and Ocean Engineering at Virginia Polytechnic Institute and State University. a Senior Member of IEEE.S. and also enables real-time dynamic modeling of the departure-prone edges of the flight envelope. in control engineering from the Technical University of Catalonia and an M.S Air Force Office of Scientific Research and was leveraged by the U. adaptive control. His research interests are in the areas of dynamics and control. She can be contacted at 223 Mechanical Engineering Building. exposes its behavior in various demanding scenarios. an M. respectively. robust control. which enables the use of arbitrarily fast estimation rates without sacrificing robustness. Urbana.D.S. adaptive and intelligent systems. Fast adaptation is critical for ensuring safe operation and providing predictable aircraft behavior for the pilot. In this sense.S. NASA’s flight tests on subscale commercial jet verify the theoretical claims in a set of safety-critical test flights. Gregory is a senior research engineer at NASA Langley Research Center. such as stall and post-stall regimes.edu) received her M. robust control. In 1997 she was awarded a governmental postdoctoral scholarship to work in INRIA. The flight test program was supported by NASA’s Aviation Safety Program. candidate in the Department of Aerospace Engineering at the University of Illinois at Urbana-Champaign. in control and dynamical systems from the California Institute of Technology in 2004. She has coauthored one book and more than 250 refereed publications. such as stall and post-stall. In 1998 she was invited to the School of Aerospace Engineering of Georgia Tech.D. ACKNOWLEDGMENTS The development of L1 adaptive-control theory was supported by U. candidate at the Department of Aerospace Engineering at the University of Illinois at Urbana-Champaign. He is the author of more than 100 publications. nonlinear systems. both in 2007. University of Illinois at Urbana-Champaign. Several architectures of the theory and benchmark examples are analyzed. in Moscow. He received an M. in theoretical mechanics and applied mathematics in 1988 from Yerevan State University in Armenia. Dynamic Systems and Control Branch. where she worked as a re102 IEEE CONTROL SYSTEMS MAGAZINE search faculty member until 2003. and a Ph. the L1 controller enables modeling of unsteady nonlinear aerodynamic phenomena at challenging flight conditions. She is an associate fellow and life member of AIAA. Prior to that.S. Her research interests are in the theory of robust adaptive control and estimation with an emphasis on aerospace applications. in aeronautics and astronautics from the Massachusetts Institute of Technology in 1988. the L1 flight control provides a reliable tool for enhancing the experimental capabilities of the GTM aircraft. Enric Xargay is a Ph. CONCLUSIONS This article presents the development of L1 adaptive-control theory and its application to safety critical FCS development.S. In 2003 she joined the Department of Aerospace and Ocean Engineering of Virginia Tech.D. His research interests include aircraft flight control. Army Research Office and U. and a member of AMS and ISDG. AUTHOR INFORMATION Naira Hovakimyan (nhovakim@illinois. She received her Ph. Irene M. respectively. and M. multiagent systems. in 1990. He earned his B. Evgeny Kharisov is a Ph.D. from the Institute of Applied Mathematics of Russian Academy of Sciences. real-world environment that stresses the control law to its limits. Rohrs’s example and the two-cart system are used as benchmark problems for illustration.S. and in 2008 she moved to the University of Illinois at Urbana-Champaign. where she is a professor and Schaller faculty scholar. in mechanical engineering from the Massachusetts Institute of Technology in 2004. Chengyu Cao has been an assistant professor of mechanical engineering at the University of Connecticut since 2008. The key feature of L1 adaptive-control architectures is the decoupling of estimation and control. even as significant changes in aircraft dynamics happen unexpectedly. His research interests include adaptive control. degree in electronics and information engineering from Xi’an Jiaotong University. Office of Naval Research. He received his Ph. China. and automatic control in aerospace system. The authors would like to thank the anonymous reviewers whose comments helped to improve the presentation. with a focus on unmanned systems and aerospace applications. He received the bachelor’s and master’s from the Department of Informatics and Control Systems at Bauman Moscow State Technical University in 2004 and 2007. She is the 2011 recipient of the AIAA Mechanics and Control of Flight Award.S.Flight Test Summary The series of flight tests on the GTM remotely piloted aircraft conducted with an L1 adaptive-control law provides a challenging. Moreover.

Las Vegas.. and N.” in Proc. Englewood Cliffs. 3. 1984. Philadelphia.. Mediterranean Conf.” in Proc. “Global stability of parameter-adaptive control systems. subsonic flight regimes. 27. B. pp. Ioannou and P. 1958. no. V. Valavani. V. Hovakimyan. Ioannou and J. 1–20. NJ: Prentice-Hall. [25] A. Decision and Control. New York: Dover. and N. Automat.. Navigation and Control Conf. V. 2005. “Adaptive control and the X-15. 1985. Adkins. Portland. Landau. 2010. 4. [26] K. Annaswamy. Cao. Keystone. Toronto. [9] P. vol. 1169–1175. 468–478. Hovakimyan. J. [8] I.” in Proc. Princeton University Conf. Wise. Cao and N. vol. J. vol. AIAA-2006-6426. no. 1979. 1982. Ioannou and P. Aircraft Flying Qualities. San Francisco. M. 6. “Closing certification gaps in adaptive flight control software. 2. [42] C. J. E. pp. Automat. “Stabilization of cascaded systems via L1 adaptive controller with application to a UAV path following problem and flight test results.” in Proc. E. 921–927. Hovakimyan and C. 30. Englewood Cliffs. NJ: Prentice-Hall. NJ: Prentice-Hall. OCTOBER 2011 « IEEE CONTROL SYSTEMS MAGAZINE 103 . K. and Control (Interdisciplinary Applied Mathematics). S. NJ: Prentice-Hall. Taylor. and Adaptive Control. Adaptive Control Processes—A Guided Tour. “Failures of adaptive control theory and their resolution. Narendra and A. vol. “Benchmark problems of adaptive control revisited by L1 adaptive control. New York: Springer-Verlag. Sastry and M. Varaiya. 2010. [2] H. Stochastic Systems: Estimation. “Adaptive control around 1960. [15] K. M. and Control Conf. CA. [31] P. “Adaptive flight control of a sensor guided munition. and J. 1984. Contr. no.. Narendra. Hovakimyan. M. Upper Saddle River. and L. and E. New York: Marcel Dekker. “An L1 adaptive pitch controller for miniature air vehicles. 34th IEEE Conf. W. Automat. Hovakimyan. [33] B. “A new adaptive law for robust adaptation without persistent excitation.” IEEE Trans. vol. 6. AIAA-20066777. 2005. D. pp. Honolulu. Minneapolis. Mishkin and L. Åström. Automat. N. W. vol. Adaptive Control. Contr. and E. 1980. 1982.. Egardt. pp. B. Cao. Contr. [12] L. [13] L. no. Xargay. 32. and V. CO. [23] R. Englewood Cliffs. AIAA2005-6385. no." in Proc.-H. 2. Åström. A. “Stable model reference adaptive control in the presence of bounded disturbances. NJ. 80. 1959. Ljung. [14] R. 2006. pp. [16] B.. J. Contr. I. Matthews. pp. pp.. and G. Navigation. 433–439.” IEEE Trans. 1973. Sin. and N. pp. C. C.. Identification. Li. N. vol. A. pp. 3. “Stability margins of L1 adaptive control architecture. P. [38] K. REFERENCES [1] S. “Flight test of an L1 adaptive control law: Offset landings and large flight envelope modeling work. Wittenmark. Syst. no.flight control for hypersonic. AIAA Guidance. 2008. 1965. pp. 55. “Robust redesign of adaptive control. vol. Xargay. Navigation and Control Conf. Self Adaptive Flight Control Symp. [17] G. Kharisov. “Air research and development command plans and programs. [24] Y. D. 29. Adaptive Control: Stability. N. E. [20] J. Peterson and K. Patel. “On self-tuning regulators. A. Lavretsky. Nonlinear Systems. [29] B. Narendra. Lin. Robust Adaptive Control. “Application of L1 adaptive controller to wing rock. NJ: Prentice-Hall.” IEEE Trans. [3] C.” IEEE Trans. 1276–1281. Kaminer. [6] E. [37] K. S. AIAA Guidance. Gregory. C. Contr. C.” Automatica. N. 185–199. no. 2007. CO. System Identification—Theory for the User. S. Hovakimyan. Tsakalis.” IEEE Trans. She is an Associate Fellow of AIAA and a Member of IEEE. 1989. Jacklin. Automat. 2006. 2006. New York: McGrawHill.. 1961. “Performance limitations of adaptive parameter estimation and system identification algorithms in the absence of excitation. [19] S. S. pp. Dixon. Adaptive Filtering Prediction and Control.” in Proc. Contr. “Design of self-optimizing control systems. “Bounded error adaptive control. 5. Francis. NJ: Prentice-Hall. J. 202–211. [11] K. AIAA2010-8047. Stability of Adaptive Controllers.” in Proc.. Goodwin and K. C. 5966–5971. Whitehead.. Keystone. 1787–1792. 1991. pp. nonlinear control. Åström and B. Gregory. 1996. 1983. IEEE Conf. [32] P. supersonic. 2010. and B. “Comparison of several adaptive controllers according to their robustness metrics.” Automatica. vol. Sun. [39] K. 1982. Anderson. 549–560. [30] P. “Interactions between excitation and unmodeled dynamics in adaptive control. New York. O.. Englewood Cliffs. Hovakimyan.” in Proc.” IEEE Trans. pp. Hovakimyan. Automat. vol. AIAA Guidance. and C. Khalil. AIAA-2008-6988. Jr. Contr. Inform. their robustness and performance measures for real systems as well as autonomy and safe flight in next generation national aerospace system. Dobrokhodov. E. Braun. Navigation and Control Conf. 2010. Bodson. AIAA Guidance. Sastry. Control and Automation. 4. vol. J. Adaptive Control: The Model Reference Approach (Control and Systems Theory). Athans.” in Proc. S.” in Proc. 1995. E. Englewood Cliffs. Her current interests are adaptive. L1 Adaptive Control Theory. 1987. 2009. Originally published by Addison Wesley. [7] I. N. Kokotovi´ c. 881– 889. “Flight test of L1 adaptive controller on the NASA AirSTAR flight test vehicle. launch vehicles. “Adaptive control in flight: Theory.” Commun. 1999. A. R. “An asymptotic error analysis of identifiers and adaptive observers in the presence of parasitics. [18] S. 1984. pp.” in Proc. [35] K. Lavretsky. no. Adaptive Control Systems. HI.. NJ: Springer-Verlag.” in Proc. 1987. [5] E. Wise. J. no. AIAA-2011-6608. Knoebel. M. 2011. L. “Robustness of continuous-time adaptive control algorithms in the presence of unmodeled dynamics. Narendra. vol. [21] P. Toronto. Gregory. Beard. [34] G. D.. 2. Automat. Cao. [10] E. American Control Conf. Convergence and Robustness. Åström. Contr. 440–448. Cao. 31–36. S. Englewood Cliffs. 1979. [27] C. 2nd ed. [40] R. Ioannou and P. NJ: Princeton Univ. application.” in Proc. 1961. Hovakimyan. 32. 1. 2008. 3. pp. no. pp. AIAA Guidance. Y. [4] N. Rohrs. Press. American Control Conf. AIAA-20108015. Adaptive Systems with Reduced Models. and K. 2784–2789. Applied Nonlinear Control. Navigation and Control Conf. pp. 1980. Kumar and P. Lavretsky. A. Gregory. A. Nonlinear Systems: Analysis. 9. 25.” IEEE Trans. A. 1995. Zimmerman. [22] K. Stability. 1996.. Automat. Greece. Decision and Control. Automat.” ASME Trans. V. Slotine and W. 2002.. OH: Wright-Patterson Air Force Base.-J. S. P. no. NJ: Prentice-Hall. 25. part II: Proof of stability. Valavani. Cao. Ed.” in Proc. 8–15. E. Bellman. J. Princeton. S. 9. S. New Orleans. Kalman. 27. Morse. Navigation and Control Conf. and open problems. S. Kokotovi´ c. 27. Navigation and Control Conf. 134–145. 480–487. AIAA Guidance. [41] C. Åström and B. Kresselmeier and K. Kokotovi´ c.” IEEE Trans. Xargay. pp. 1161–1168. Cao. 1986. [28] K.. no. pp. C. aircraft. vol. Secaucus. N. MN. and unmanned aerial vehicles. LA. V. OR. Princeton. New York: Springer-Verlag. [36] P.. “Stable adaptive controller design. PA: Society for Industrial and Applied Mathematics. Hovakimyan. E.. Canada.” IEEE Trans. NV. Wittenmark.. Cao.. Stein. Contr. AIAA Guidance. Thessaloniki.

[49] I. A. no. Lizárraga. Cao. “L1 adaptive neural network controller for autonomous aerial refueling with guaranteed transient performance. Gato. AIAA-2008-6987. Li. Guid.” in Proc.. C. 2009. 53. Hovakimyan. Åström and B. Hovakimyan. Hovakimyan.cfm?ident˙number=7003 7&StartRow=1&PaginatorPageNum ber=1&doc_id=MIL–STD–1797&search_method=BASIC [73] B. [55] T. Hovakimyan.assistdocs.. [68] E. Adaptive Control. 1997. Available: http://www. Hilton Head. 1200–1221. Hovakimyan. 10. E. 729–740.” AIAA J. Pascoal. [45] J. 2009. V.” Int. M. and E. no. Navigation and Control Conf.S. and X. IL. St. C. N. stable L1 adaptive output feedback controller. 33. C. “Model-based L1 adaptive control of hysteresis in smart materials. 2006. 2009. MA: Addison-Wesley. and K. Hindman. K. O. Cao and N.” in Proc. and G. 2009. Cunningham. Gregory.” IEEE Trans. Lavretsky. M. Cao. Lei. 2007. 5272–5277. A. Praly. 37. E. Kaminer. Hovakimya n. Louis. I. Navigation and Control Conf. MD. AIAA Guidance. vol. AIAA Guidance. Wang. [52] V. Wise. C.” in Proc. N. vol. Patel. Cao. vol. Navigation and Control Conf. Hovakimyan. Chicago.. and V. Kaminer. American Control Conf.” in Proc. M. 2007. “A piloted evaluation of damage accommodating flight control using a remotely piloted vehicle. and N. Gregory. Cox. 2. Wittenmark. “Coordinated path following for time-critical missions of multiple UAVs via L1 adaptive output feedback controllers.” in Proc. 32. Lavretsky. Yakimenko. N. “Path following for unmanned aerial vehicles using L1 adaptive augmentation of commercial autopilots. L. J. Navigation and Control Conf. M. [77] J. “L1 adaptive control design for NASA AirSTAR flight test vehicle. Hovakimyan. Wang. Hovakimyan.. M. Lavretsky. New York. Lei. 3–11. Dobrokhodov. Dyn. C. C. Kaminer. Pomet and L. CA. Xargay. V. Rivers.” in Proc.” in Proc. Wise. Lewis. Hovakimyan.” J. Hovakimyan. [61] C. Guid. Jones. Dobrokhodov. Balas. Cancun. Stewart.” in Proc. Honolulu. N. AIAA-2007-6644. B.. C. Pascoal.. progress and new results in robust adaptive control. [67] E. C.. [64] J. D. Hovakimyan. Kharisov and N. Zou. K. 2007. V. M. 2008.. N. Hovakimyan. Hovakimyan.” in Proc. I. Xargay. AIAA Guidance. 3. P. Fan and R.. Dobrokhodov. 3251–3256. How. AIAA Guidance. C. Cao. vol. I. Young. Control Dyn. V. [71] G.” in Proc. Navigation and Control Conf. Chicago. pp. SC. C. Patel. Jr. Dobrokhodov. Navigation. Hilton Head. vol. Wilborn.... Navigation and Control. Rohrs. IL. A. Smit h. Dorobantu. CA. “Designing a high performance. N. 2008.” in Proc. E. pp. Gregory. V. SC. AIAA Guidance. 2007. “Adaptive nonlinear regulation: Estimation from the Lyapunov equation. “Issues. Fremaux. and S. E. Navigation and Control Conf..” in Proc. “L1 adaptive controller for a missile longitudinal autopilot design. Shah. “Dynamics modeling and simulation of large transport airplanes in upset conditions. [74] E. V. Foster. and C. [44] V. E. Control Signal Process. Hovakimyan. AIAA-2009-5619. AIAA Guidance. Kitsios. Xargay. [78] K. pp. AIAA-2009-5681. AIAA Guidance. HI. AIAA-2008-7128.. V. and N. 1182–1187. Boston. [58] B. Automat. Navigation and Control Conf. and A. V. 1994. G. Navigation and Control Conf.-B. “Robustness of adaptive control algorithms in the presence of unmodeled dynamics.” IEEE Trans. [70] J. Cao. 42. IEEE Conf. Kaminer. Baltimore. 586–591. M. C. “Adaptive control laws for flexible semi-span wind tunnel model of high-aspect ratio flying wing. S. Hovakimyan.” in Proc. B. 2008. AIAA Guidance. [63] C. G. Aug. Automat. and G. Hovakimyan. 2005 AIAA-2005-533. D. 20. [72] MIL-STD-1797B. Hilton Head. R. Kharisov. “Flight validation of metrics driven L1 adaptive control. and K. [66] C. E. Kaminer. Guidance. 2011. Krishnakumar. V. Decision and Control. Patel. E. Valavani. vol. A. M. Pascoal. Navigation and Control Conf. Cao. and E.” in Proc. Cliff. C. AIAA-2009-5754. Canada. Toronto.[43] V. Xargay. Cao. Athans. M.” in Proc. Control. SC. “L1 adaptive output feedback controller for minimum phase systems. 2007.com/search/document_details. Leman. 1992. MO. Honolulu.” in Proc. AIAA-2009-6188. J. M. and D. and N. Xargay. [Online]. and N. San Francisco. N. Portland. no. AIAA Guidance. CEAS Conf. Navigation and Control Conf. HI. and E. and I. Cao. [59] I. OR. Georgiou and M. 2010. pp.” in Proc. “Robustness analysis of nonlinear feedback systems: An input-output approach. FL. G.. J. N. Cao. 2009. Navigation and Control Conf. New York: Wiley. [65] T. [69] S. IL. Orlando. Patel.. pp. Hovakimyan. Seiler. AIAA-2009-5738. 874–879. [54] I. Cao. 550–564. Note D–5153. Contr. Honolulu. San Francisco. AIAA-2010-8407. A. [46] J. 2010. “L1 adaptive neural network controller for autonomous aerial refueling in the presence of unknown actuator failures. Cao. Wang. Interface Standard (Limited circulation).” in Proc. American Control Conf. American Control Conf. I.. I. Michini and J. M. 2007. Gregory. E. » OCTOBER 2011 . AIAA Guidance. E. “Design and analysis of a novel L1 adaptive control architecture with guaranteed transient performance.. and A. 1969. CA: Ames Research Center. [50] Y. Navigation and Control Conf. HI. R. “ L1 adaptive controller for air-breathing hypersonic vehicle with flexible body dynamics. 2009. Fekri. 2008. Cliff.” in Proc. Navigation and Control Conf. Cao. no. Stein. C. C. K.. Cooper and R. Hovakimyan. SC. Germany. H. Cunningham. V. and N. L.. 2010. E. “ L1 adaptive control augmentation system for the X-48B aircraft. Cao and N. Decision and Control. C.” in Proc. Flying qualities of piloted aircraft. Navigation and Control Conf. Hindman. 1982. “L1 adaptive control law in support of large flight envelope modeling work. “The use of pilot rating in the evaluation of aircraft handling qualities. Cao. “L1 adaptive control for indoor autonomous vehicles: Design process and flight testing. Kitsios. Riddick. Dullerud. IEEE Conf. A. Cao. E. AIAA Guidance. AIAA-2007-6525. Hilton Head. pp. Cao. [48] R.. S. C. Ridgely. pp. Patel. AIAA Guidance. 3166–3171. 519–579. C. N.. 2009. Lizárraga. [76] I.” in Proc. Kurdila. Hilton Head. I. pp. “Robustness analysis of an L1 adaptive controller. “Experimental validation of a metrics driven L1 adaptive control in the presence of generalized unmodeled dynamics. and R. L.. “Design of an L1 adaptive controller for air-breathing hypersonic vehicle model in the presence of unmodeled dynamics. AIAA Guidance. and I.” in Proc. AIAA-2010-7773. SC. 6. C. A. N. Canada. C.. E. Jones. Athans. Mexico. “L1 adaptive controller for tailless unstable aircraft. I. V. [53] I. C. Gregory. Navigation and Control Conf. AIAA-2006-6206. Moffett Field. Hovakimyan. K. V. AIAA-2007-6409. (2006). Dobrokhodov. Tech. “ L1 adaptive output-feedback controller for non-strictly-positive-real reference systems: Missile longitudinal autopilot design. American Control Conf. Chicago. E. Kaminer. N. AIAA Guidance. Navigation and Control Conf. J. N. M. Harper. AIAA-2006-6314. Chicago. “Development and implementation of L1 gimbal tracking loop onboard of small UAV. Dobrokhodov. AIAA-2006-6527. and Control Conf.” in Proc. Hovakimyan. 2011. [75] P. Aircraft Control and Simulation. vol. no. 2010. V. no. AIAA-2006-6313.. 1992. 2011. pp. Adapt. Gregory. M. “L1 adaptive control for flexible space launch vehicle and proposed plan for flight validation. J. Navigation and Control Conf. IL. N. and E. AIAA Guidance. SC. 2008.” NASA. N. Keystone. Statnikov. Hovakimyan. Department of Defense Military Specification.” IEEE Trans. T. [56] Z. 9.. I.” in Proc. 717–726. C. D. E.” in Proc. T. Cao. N. 2. and A. Lavretsky. A. “L1 adaptive controller for tailless unstable aircraft in the presence of unknown actuator failures. V. Hilton Head.” in Proc. Patel. pp. [62] K. Xargay. E. Kurdila.. M. [47] I. AIAA Guidance. Wise. Stevens and F. Cao. CO. N. Chicago. M. V. K. Nguyen. Contr. Hovakimyan. Gregory... I. 104 IEEE CONTROL SYSTEMS MAGAZINE [60] X. Murri. and W. AIAA Guidance. “L1 adaptive flight control system: Systematic design and verification and validation of control metrics. Smith. pp. AIAA Guidance. AIAA-2011-6451. “L1 adaptive controller for multiinput multi-output systems in the presence of nonlinear unmatched uncertainties. 2006. Toronto. 1. Contr. [51] E. D. Automat. J. pp. IL. Xargay. M. 2007. U. A. Navigation and Control Conf. [57] Y. AIAA-20086282. AIAA Guidance. Cao. Xargay. Xargay. Munich.