You are on page 1of 10

Applied Soft Computing 9 (2009) 553–562

Contents lists available at ScienceDirect

Applied Soft Computing
journal homepage: www.elsevier.com/locate/asoc

Controlling a drone: Comparison between a based model method
and a fuzzy inference system
K.M. Zemalache *, H. Maaref
IBISC Laboratory, CNRS-FRE 3190, Universite´ d’Evry Val d’Essonne, 40 Rue du Pelvoux, 91020 Evry Cedex, France

A R T I C L E I N F O

A B S T R A C T

Article history:
Received 2 October 2007
Received in revised form 23 July 2008
Accepted 10 August 2008
Available online 11 September 2008

The work describes an automatically on-line self-tunable fuzzy inference system (STFIS) of a new
configuration of mini-flying called XSF (X4 Stationnary Flyer) drone. A fuzzy controller based on on-line
optimization of a zero order Takagi–Sugeno fuzzy inference system (FIS) by a back propagation-like
algorithm is successfully applied. It is used to minimize a cost function that is made up of a quadratic
error term and a weight decay term that prevents an excessive growth of parameters. Thus, we carried
out control for the continuation of simple trajectories such as the follow-up of straight lines, and complex
(half circle, corner, and helicoidal) by using the STFIS technique. This permits to prove the effectiveness of
the proposed control law. Simulation results and a comparison with a static feedback linearization
controller (SFL) are presented and discussed. We studied the robustness of the two controllers used in the
presence of disturbances. We presented two types of disturbances, the case of a breakdown of an engine
as well as a gust of wind.
Published by Elsevier B.V.

Keywords:
Self-tunable fuzzy inference system
Static feedback linearization controller
Tracking control
Dynamic systems
Drone

The past few years have witnessed a rapid growth in the use of
fuzzy logic controllers for the control of processes, which are
complex and ill defined. Most fuzzy controllers developed till now
have been of the rule-based type, where the rules in the controller
attempt to model the operator’s response to particular process
situation.
Recently, the resurgence of interest in the field of artificial
neural networks has injected a new driving force into the fuzzy
literature. The back-propagation learning rule, which drew little
attention till its applications to artificial neural networks was
discovered, is actually an universal learning paradigm for any
smooth parameterized models, including fuzzy inference systems.
As a result, a fuzzy inference system can now not only take
linguistic information from human experts, but also adapt itself
using numerical data (input/output pairs) to achieve better
performance. This gives fuzzy inference systems an edge over
neural networks, which cannot take linguistic information directly.
In autonomous wheeled robot, many developed learning
techniques have arisen in order to generate or to tune fuzzy rules.

Most of them are based on the so-called ‘‘neuro-fuzzy learning
algorithms’’ as proposed by [22,23,25,26,32,39]. These methods
are well for constructing an optimal fuzzy system model which is
used to identify the corresponding practical system.
Modeling and controlling aerial vehicles (blimps and mini
rotorcraft) are the principal preoccupation of the IBISC-group.
Within this optic, who attracted the contest of the DGA-ONERA1
was the XSF project which consists of a drone with revolving
aerofoils [4](see Fig. 1 left). It is equipped with four rotors where
two are directionals, what we call in the following X4 Stationary
Flyer (XSF).
The XSF is an engine of 68 cm  68 cm of total size and not
exceed 2 kg in mass. It is designed in a cross form and made of
carbon fiber. Each tip of the cross has a rotor including an electric
brushless motor, a speed controller and a two-blade ducted
propeller. The operating principle of the XSF can be presented thus:
two rotors turn clockwise, and the two other rotors turn
counterclockwise to maintain the total equilibrium in yaw motion.
The equilibrium of angular velocities of all rotors done, the
unmanned aerial vehicle (UAV) is either in stationary position, or
moving vertically. A characteristic of the XSF compared to the

* Corresponding author.
E-mail addresses: zemalache@iup.univ-evry.fr,
kadda.zemalache@ibisc.univ-evry.fr (K.M. Zemalache),
hichem.maaref@ibisc.univ-evry.fr (H. Maaref).

1
This work is supported by the mini-flyer competition program organized by the
DGA (Direction Ge´ne´rale des Armements) and the ONERA (Office Nationale d’Etude
et de Recherche en Ae´rospatiale), France.

1. Introduction

1568-4946/$ – see front matter . Published by Elsevier B.V.
doi:10.1016/j.asoc.2008.08.007

performances and stability of the suggested controller are analyzed through simulations carried out on the model (kinematics and dynamic equations). Configuration description and modeling Unlike regular helicopters that have variable pitch angles. which utilizes double cameras. The proposed controller is based upon the compensation of the Coriolis and gyroscopic torques and the use of a PD2 feedback structure. hovering and landing control of a four rotors by synthesizing a controller using the Lagrangian model based on the Lyapunov analysis. [30] proposed an attempt to apply linear H1 outer control of helicopter quadrotor with plant uncertainty combined with a robust feedback linearization inner controller. a STFIS control strategy is developed based on the systems output measures is implemented. In this paper. Also. 1(right). The robustness of the proposed controller is then evaluated in Section 5. achieved the formation control by sliding mode technique and focused on collision and obstacle avoidance by extracting the state variables with a Kalman filter. [20. Waslander et al.554 K. Motion planning and simulation results are introduced in Section 4.33. [5]. Lateral motion. intended to carry out in an autonomous way a reconnaissance mission. 2. Zemalache. is adapted and modified for the used with the XSF. Maaref / Applied Soft Computing 9 (2009) 553–562 Fig. 1. [18]. Castillo et al. H. based on Newton formalism. conclusion and future works are given in the last section. Tayebi and McGilvray [37] proposed a new quaternion-based feedback control scheme for exponential attitude stabilization of a quadrotor.M. where the proportional action is in terms of the vector quaternion and the two derivative actions are in terms of the airframe angular velocity and the vector quaternion velocity. [16] where the dynamic motor effects are incorporated and a bound of perturbing errors was obtained for the coupled system. 1 right).10. In addition to the military applications. and used an exact linearization and non-interacting control for the global system to evaluate translational motion and yaw angle outputs. the angles and the angular speeds are available for measurement. addressed the classical problem of speed estimation of an Unmanned Aerial Vehicle when the acceleration. A solution has been provided for a class of systems via the tools of adaptive observation theory with promissing results. presented a simple closed loop equipped with MEMS (microelectro-mechanical systems) sensors and PIC based processing unit. Conceptual form of the XSF (left) and frames attached to the XSF (right). [19]. This permits a more stabilized horizontal flight and a suitable cornering. which exploits the Moir patterns.2. related models for controlling the vertical take-off and landing (VTOL) aircraft are studied by Hauser et al. Mokhtari et al. [40] maked an emphasis on the insufficiency of classical control methods and proposes the integral sliding mode controller associated with reinforcement learning to achieve multi agent control.42]. existing quadrotors. [28] developed a non-linear controller with observers based on the backstepping.41. was studied by Hamel et al. and in [3].21]. the environmental protection. Mederreg et al. This swiveling ensures either the horizontal rectilinear motion or the rotational movement around the yaw axis or a combination of these two movements which gives the turn (see Fig. Vertical motion is controlled by collectively increasing or decreasing the power for all motors. The role of the trajectory generator is to generate a feasible time trajectory for the UAV. Bestaoui and Slim [8] addressed the problem of characterizing maneuvers paths on the group of rigid body motions in 3D for a quadrotor. The robustness of this controller is studied. In Benzemrane et al. A model for the dynamic and configuration stabilization of quasi-stationary flight conditions of a four rotors VTOL. Hamel and Mahony [17] proposed a vision based controller which performs visual servo control by positioning a camera onto a fixed target for the hovering of a quadrotor. an engine has fixed pitch angle rotors and the rotor speeds are controlled to produce the desired lift forces. in particular a camera. Finally. Flight control methods utilizing vision systems are also studied by [38]. The developed ideas of control for the XSF by the self-tunable fuzzy inference system (STFIS) controller is presented and compared with a static feedback linearization controller (SFL) to stabilize the XSF by using the point to point steering stabilization in Section 3. is not achieved by differentially controlling the motors generating a pitching/rolling motion of the airframe that inclines the collective thrust (producing horizontal forces) and leads to lateral accelerations . The stabilization problem of a four rotors rotorcraft is also studied and tested by Castillo [12] where the nested saturation algorithm is used. the natural risk management. in particular the search for anybody in buildings in flame. Drone XSF is a flying machine of low dimension able to carry a small payload. This technique early used for autonomous wheeled robot. Several recent works were completed for the design and control in pilotless aerial vehicles domain such that Quadrotor [1. in [11] a proportional integral derivative (PID) controller and a linear quadratic (LQ) controller were implemented and proved capable of regulating the system and application of the theory of flat systems by Beji and Abichou [6]. the management of the great infrastructures. the input/output linearization procedure [18]. [13] performed autonomous take-off. The arrangement of this paper is as follows. Hanford et al. Hoffman et al. [31] developed a dynamic model in non-linear state space representation. this type of machines can also interest the civil field. Mistler et al. All the reviewed techniques require the well knowledge of the system dynamic model and parameters. in x direction or in y direction.29. The dynamic model of the XSF drone is given in the second section. Basic motions of the XSF can described using Fig. is the swiveling of the supports of the motors 1 and 3 around the pitching axis thanks to two small servomotors controlling the swiveling angles j1 and j3 .

Iyy and Izz are of the inertia matrix IG expressed in G. Izz Þ.16]. 1(right)). This swiveling ensures either the horizontal rectilinear motion or the rotational movement around the yaw axis or a combination of these two movements which gives the turn (see Fig. while the second time derivative permits to ðdOGÞ=ðdtÞ ¼ ðx˙ y˙ zÞ 2 ¨ T (Fig. Ey . Eg1 . Ez g such that the vertical direction Ez axis is pointing upwards. The aim is to control the engine vertically z axis and horizontally according to x and y axis. 0 1 C u Sc Su CcCu (1) R ¼ @ Sf C c Su  Sc C f Su Sc Sf þ C c C f C u Sf A Su C c C f þ Sc Sf C f Su Sc  C c Sf C u C f where for example C u and Su represent cos u and sin u. The vehicle dynamics. The XSF is a system consisting of four individual electrical fans attached to a rigid cross frame. Let 0 1 0 1 0 1 0 0 0 e´ 1 ¼ @ Sj1 A . Eg2 . t c ÞT as an auxiliary inputs. Drag forces and gyroscopic due to motors effects will be not considered in this work. is modeled by the system of following [4.u. the model given above can be input/output linearized by the following decoupling feedback laws t u ¼ Ixx Sf f˙ u˙ þ Ixx C f t˜ u 2 t f ¼ Iyy Sf C u f˙  Iyy Su C f u˙ f˙ þ Iyy C u C f t˜ f t c ¼ Izz t˜ c (10) and the decoupled dynamic model of rotation can be written as h¨ ¼ t˜ (11) with t˜ ¼ ðt˜ u t˜ f t˜ c ÞT Using the system of equations (2) and (11).12. Zemalache. we get (5) j1 and j3 are the two internal degree of freedom of rotors 1 and 3. two engines of direction are used to permute between the x and y motion. e2 ¼ e4 ¼ @ 0 A (4) C j1 R C j3 R 1 R G G G Then we deduce: u2 ¼ f 1 Sj1 þ f 3 Sj3 u3 ¼ f 1 C j1 þ f 3 C j3 þ f 2 þ f 4 555 bounded 20  j1 .7. u (pitch). the dynamic of the system is defined by mx¨ ¼ Sc C u u2  Su u3 my¨ ¼ ðSu Sc Sf þ C c C f Þu2 þ C u Sf u3 mz¨ ¼ ðSu Sc C f  C c Sf Þu2 þ C u C f u3  mg u¨ ¼ t˜ u f¨ ¼ t˜ f c¨ ¼ t˜ c (12) The rotational part can be easily linearized with static feedback control laws. Maaref / Applied Soft Computing 9 (2009) 553–562 [1. An Euler angle representation given in Eq. 2. respectively. The lift (collective) force u3 and the direction input u2 are such that 0 1 0 @ u2 A ¼ f 1 e´ 1 þ f 2 e2 þ f 3 e´ 3 þ f 4 e4 (3) u3 with f i ¼ K T v2i where K T ¼ 105 N s2 and vi is the angular speed resulting of motor i.44]. j3  þ20 . These variables are controlled by dc-motors and u¨ ¼ u4 f¨ ¼ u5 c¨ ¼ u6 (13) . zÞ denotes the G position with respect to RO . Motion dynamic We consider the translation motion of RG with respect to RO . The constraints as external perturbation and the gyroscopic torques are neglected. Then. its time derivative gives the velocity wrt to RO such that ˙ T .16].2. The position of the mass center wrtRO is defined by OG ¼ ðx y zÞT . meaning that: h¨ ¼ P G ðhÞ1 ½t  P˙ G ðhÞh˙  (8) With t ¼ ðt u . Where the inertia elements Ixx . But. cÞ.M. We consider a local reference airframe RG ¼ fG. e2 and e4 are the unit vectors along Eg3 which imply that rotors 2 and 3 are identical of that of a classical quadrotor [43. Eg3 g at G (mass center) while the inertial frame is denoted by RO ¼ fO. 3(right)). get the acceleration ðd OGÞ=ðdt 2 Þ ¼ ðx¨ y¨ zÞ The model is a simplified one’s.1. u. The vector u2 and u3 combines the principal non-conservative forces applied to the engine airframe including forces generated by the motors and drag terms. then IG ¼ diagðIxx .u.12. as well as the direction of rotation of the rotors implies that rotors 1 and 2 turn clockwise and rotors 3 and 4 turn counterclockwise. Let the vector X ¼ ðx. H. (6) is ensured. y. And 0 P G ðhÞ ¼ @ Ixx C f 0 0 0 Iyy C f C u 0 1 0 0 A Izz (9) As a first step. Ex . Rotational motion The rotational motion of the XSF will be defined wrt to the local frame but expressed in the inertial frame. The rotation of the rigid body is defined by Rf.K. represented in Fig. 1 ðt þ Ixx Sf f˙ u˙ Þ Ixx C f u 1 2 f¨ ¼ ðt þ Iyy Sf C u f þ Iyy Su C f u˙ f˙ Þ Iyy C u C f f u¨ ¼ c¨ ¼ (6) tc Izz With the three inputs in torque t u ¼ lð f 2  f 4 Þ t f ¼ lð f 1 C j1  f 3 C j3 Þ t c ¼ lð f 1 Sj1  f 3 Sj3 Þ þ KM ð f C  f 3 C j3 þ f 4  f 2 Þ K T 1 j1 (7) where l is the distance from G to the rotor i and K M ¼ 9  106 N m s2.c : RO ! RG . f (roll) and c (yaw). where Rf. Iyy . e´ 3 ¼ @ Sj3 A .c 2 SOð3Þ is an orthogonal rotation matrix which is defined by the Euler angles. 1(right). respectively. regrouped in h ¼ ðf. (1) has been chosen [43]. 2. The equality from Eq. The main feature of the presented XSF in comparison with the existing quadrirotors [1.45]: mx¨ ¼ Sc C u u2  Su u3 my¨ ¼ ðSu Sc Sf þ C c C f Þu2 þ C u Sf u3 mz¨ ¼ ðSu Sc C f  C c C f Þu2 þ C u C f u3  mg (2) m is the total mass of the vehicle. is the swiveling of the actuators supports 1 and 3 around the axis of pitching (angles j1 and j3 ). t f .

556 K. with the static feedback laws. left) needs two STFIS networks: (1) a first networks to identify the drone (Model). (15)) and for c ¼ p=2 expression Eq.2. (15) and (19) continuity is recommended allowing the avoidance of the peak phenomenon.1. f 2 . while the rotor force-inputs are of six order F ¼ ð f 1 . j1 . the form of which is chosen through Gi j . f 4 . While we keep the input u3 for the altitude z stabilization. nx and nz can be deduced as in [7]: Remark.2. the equivalent system of controlinputs presents five inputs U ¼ ðu2 . z) dynamics which can be also decoupled by a static feedback law. The input layer receives the input signal.2. The output layer gives the crisp output value u by: Pr i¼1 ai wi u¼ P r i¼1 ai where wi are the synaptic weights between the second hidden layer and the output. f 3 . (2). Input u2 for the motion along x The aim here is to stabilize the dynamics of x with the control input u2 . As soon as we get u3 ðc ¼ 0Þ ¼ u3 ðc ¼ p=2Þ ¼ mg taking f ¼ u ¼ 0. u4 . 3.1. Only output parameters of the FIS are optimized in this work. u6 Þ. These operations can be traduced by the layer structure shown in Fig. 3. Presentation of STFIS The main idea is to implement the various parts of a fuzzy logic system in a multilayered neural network.45. So. Self-tunable fuzzy inference system The formal analogy between a fuzzy inference system and a multilayer neural network associated with optimization algorithms is used from the retro-propagation gradient algorithm have winded up in what is called a STFIS Network.M. we take u2 ðc ¼ 0Þ ¼ u2 ðc ¼ p=2Þ ¼ 0 with f ¼ u ¼ 0. connected with others by adjustable parameters. basically.1. u3 . kiz are the coefficients of a Hurwitz polynomial. Then a non-linear decoupling feedback permits to write the following decoupled linear dynamics: y¨ ¼ ny z¨ ¼ nz (17) Then we can deduce from Eq. In order to ensure this.36]. 2. Our second interest is the (x. Static feedback linearization controller 3. (16) is invertible. 3. Classical model based techniques such us the static feedback linearization and backstepping techniques have been investigated and used for stabilization with motion planning [7. (19) is considered. x¨ ¼ nx z¨ ¼ nz (20) Remark. Control input for z  ymotions We propose to control the y=z motion through the input u2 =u3 . To switch between the two controllers Eqs. We can regroup the two dynamics as        1 Su Sc Sf þ C c C f C u Sf y¨ u2 0 ¼  z¨ u3 g m Su Sc C f  C c Sf C u C f (16) For the given conditions in c and u the ð2  2Þ matrix in Eq. we use the backpropagation of the gradient to minimize the cost function. Advanced strategies of control The aim in this section is to make comparison between model based approaches and experts analysis involving fuzzy systems. For c ¼ 0 one uses (Eq. ny and nz are detailed in the following. 3. this kind of STFIS network implements a Sugeno’s type fuzzy inference system where the weights wi are the crisp conclusions of the rules. Each layer. (17) the linear controller: ny ¼ y¨ r  k1y ðy˙  y˙ r Þ  k2y ðy  yr Þ nz ¼ z¨r  k1z ðz˙  z˙ r Þ  k2z ðz  zr Þ (18) with the kiy .27]. The optimization of the parameters of the network can be done by various methods. which is done entirely on-line. The second one calculates the truth value ai of each fuzzy rule using the Min operator. . 3. (2) a second networks to control the drone (Controller). Consider ðc. Then the transformation U ! F is not a diffeomorphism. Zemalache. This can be asserted if we impose some constraints to the reference trajectories. Proposition. 3. 3. we have the following proposition. The first hidden layer calculates the membership grades ma i j of the inputs to their associated fuzzy subsets so the weights ai j allow to tune respectively the position and the width of the fuzzy subsets. With the new inputs 1 ðt þ Ixx Sf f˙ u˙ Þ Ixx C f u 2 1 u5 ¼ ðt þ Iyy Sf C u f˙ þ Iyy Su C f u˙ f˙ Þ Iyy C u C f f 1 u6 ¼ t Izz c u2 ¼ ðSc C f  Su C c Sf Þ1 ðmnx C f C u þ mðnz þ gÞSu Þ u3 ¼ ðSc C f  Su C c Sf Þ1 ðmnx ðSc Su C f  C c Sf Þ þ mðnz þ gÞSc C u Þ u4 ¼ (14) (19) The dynamic of x and z are decoupled and exponentially stable. Jordan [24] proposes the distal control method.2. So. Architecture and learning algorithm architecture In this work. As shown in Eq.1. uÞ 2   p=2. u5 . p=2½. 1 1 u2 ¼ mny C f Cc  mðnz þ gÞSf Cc 1 u3 ¼ mny ðSf Cu  C f tg c tg u Þ þ mðnz þ gÞðC f Cu1 þ Sf tg c tg u Þ (15) The dynamic of y and z are linearly decoupled and exponentially asymptotically stable with the appropriate choice of the gain controller parameters. having a specific function [15. This architecture (Fig. which is used by [34] under the name of JEAN (Jordan method Extended for Adaptive Neuro-control).2. H. we propose to generate the fuzzy control rules by an optimization method. Maaref / Applied Soft Computing 9 (2009) 553–562 with Proposition. j3 Þ.1. In this work.

M. some similar performances are obtained in terms of generalization mean error. 3(right) for the control of the XSF. But. ban1 wnij ðt  1Þ j P n1 j aj (23) . it modifies the itinerary taken by the learning algorithm in the action space: it will not follow the higher slope but it will stay on a downward path [24]. It uses only one network as a controller. The simultaneity of the two phases can lead to serious problems of instability for JEAN architecture. The tuning of the synaptic weights is done by back-propagation of the quantity du through the network controller. Maaref / Applied Soft Computing 9 (2009) 553–562 557 Fig. 3(right)). Optimization of adjustable parameters is accomplished with a retro-propagation gradient algorithm which is adapted to the net structure (Fig. JEAN learning architecture (left) and control architecture mini-JEAN (right). where the learning is done directly by the back-propagation of the output error. The value of du itself is obtained by back-propagation of the cost function based on the output error e (e ¼ y  y1 . we choose an architecture ‘‘mini-JEAN’’ as illustrated in Fig. 2. The aim is to minimize cost function E: E ¼ 12e2 (21) where e is the difference between set point and process output. an imprecise model of the plant does not alter the minimization of the cost function. In fact. Compared to JEAN architecture. The technique using only the approximation of the Jacobian by its sign (easier to know then the Jacobian itself) is a viable approach to control simple (SISO) plants [35]. H. it is not often necessary to know accurately the plant model (i. Self-tunable fuzzy inference system.K.e. On the other hand. where y1 is the effective output value and y is the desired output) through the network model. Zemalache. The basic equations of the algorithm are: wnij ðt þ 1Þ ¼ wnij ðtÞ þ Dwnij ðtÞ (22) With Dwnij ðtÞ ¼ bDwnij ðt  1Þ  hdni an1  j Fig. Then. the computing time favors clearly mini-JEAN. This architecture does not require an emulator network. Nevertheless. the Jacobian). to have a suitable control. 3.

84 1. H. A fuzzy controller based on an on-line optimization of a zero order Takagi–Sugeno fuzzy inference system is successfully applied. if the learning is prolonged.82 Z 7.72 39.22 39. b is the n moment parameter. a j is the truth value of the premise part of the triggered rule. Zemalache.73 41. di is the error term (ith neurone of layer n). The cost function is given by X J ¼Eþl w2i (24) Table 1 Weight for z displacement de ne NB NS NB NS Z PS PB 29.2. t is the training iteration.71 31. It is used to minimize a cost function that is made up of a quadratic error term and a weight decay term that prevents an excessive growth of parameters of the consequent part.02 4.63 19. 4.54 22. the parameters are modify as   @J wðt þ 1Þ ¼ wðtÞ þ bDwðtÞ þ h (25) @w This algorithm easily includes the effect of the second term of the cost function J and by taking b ¼ 2lh (regression coefficient) we obtain:   @E bawðtÞ  P (26) wðt þ 1Þ ¼ wðtÞ þ bDwðtÞ þ h @w a Since a fuzzy inference system is concerned. The technical characteristics of this flying vehicle were presented in [4].52 34. the quality of the control deteriorates. we adapt this formula by multiplying b by the firing term of the rule.3. The engine acquires by its systems output measures. We have considered a total mass equal to m ¼ 2 kg. algorithm control parameters and learning duration (number of iterations).69 2.62 quadratic error and the quality of the control is maintained even under prolonged learning.55 3. Algorithm modification weight regression The procedure is entirely done on-line on the engine. the weights increase continuously with time and progressively.99 1. calculates the cost function to back-propagation. If we limit the optimization only on the conclusions parameters w41 j . Fig. Starting with a preinitialized rules table.79 9.27 PS PB 2. 3.92 29.62 29. The second term in J is known as weight decay and used usually in the context of classification problems.37 39.34 29. The weights of the table of decision are then adjusted locally and progressively.59 10. the authors have noticed. it performs the acquisition of the distances (observations). By applying this method to our system. The used method is based on a back-propagation algorithm where the parameters values are free to vary during the optimization process. the reader is referred to the web version of the article. This technique early used for autonomous wheeled robot [26] but. Maaref / Applied Soft Computing 9 (2009) 553–562 where wnij ðtÞ is the ith parameter between i of layer n and jth unit of layer n  1. The main idea is to generate the conclusion parts (so-called weight) of the rules automatically thanks to an optimization technique. The quality of solution obtained using this algorithm depends on input learning signals. begins to move and so on.96 2. w are the parameters (weights) to optimize parameters and l is a constant that controls the growth of parameters.43] for more details. The weights wi are then adjusted locally and progressively.61 9. Then. we use here a well known technique in the field of neural classification methods [14] and having a strong relationship with ridge regression and regularization theory [9]. We solve the tracking control problem using the point to point steering stabilization see [7.M. Motion planning and simulation results The XSF is tested in simulation in order to validate some motion planning algorithm considering the proposed STFIS control laws.32 28.558 K. we get Dw41 j ðtÞ ¼ bDw41 j ðt  1Þ  hd41 a3j  ba3j w41 j ðt  1Þ P 3 j aj (27) with where E is the classic quadratic error. n1 a j is the output of jth unit of layer n  1 and b is the regression coefficient. a saturation of weights growth is obtained without any degradation of the residual d41 ¼ P e j a3j (28) 4. when XSF begins to fly. calculates the error to the back-propagated. updates the triggered rules in real time. h is the learning gain.) . To overcome this difficulty. The shape of the used membership functions is triangular and fixed in order to extract and represent the knowledge from the final results easily.74 7. (For interpretation of the references to colour in the artwork. Evolution of the weights for z displacement (left) and representation of the linguistic translation of the controller u3 (right). namely P a j = a j . Thanks to the classic back-propagation algorithm. The table of rules (weights wi ) can be initially empty or filled with an a priori knowledge. updates the triggered rules on-line.

Realization of a round corners (left) and inputs u3 and u2 for the realization of a round corners (right). we use the MIN operator for the composition of the input variables. The use of a function cost gathering a quadratic error and a term of regression of the weights enabled us to achieve our goal. 5] Very Weak. the reader is referred to the web version of the article. The universes of discourse are normalized and shared in five fuzzy subsets for all displacement. The linguistic labels are defined as follows: NB: Negative Big. we can claim that the extracted rules are quite logical and coherent. 5 and 6 (right) illustrates the controlled positions zxy using STFIS controller where u3 and u2 . (For interpretation of the references to colour in the artwork.) Fig. PS: Positive Small and PB: Positive Big. Z: approximately Zero. So. On the other hand. W: [7. Note that the input u3 ¼ mg at the equilibrium state is always verified. 11] Weak. we can observe that the two sets of linguistic rules are quite close. By comparing the table proposed by learning and by human expertise (see Tables 2 and 3). Maaref / Applied Soft Computing 9 (2009) 553–562 Table 2 Learning linguistic table de ne NB NS Z PS PB NB NS Z PS PB B B VB VB VB B B B VB VB W W M M* B VW VW VW* VW* W VW VW VW W W Table 3 Expertise linguistic table de ne NB NS Z PS PB NB NS Z PS PB B B VB VB VB B B B VB VB W W M B B VW VW W W W VW VW VW W W The deduce the truth value. 42] Very Big. Three cases (noted with *) are different and they differ from only one linguistic concept (M instead B and VW instead W). Zemalache. Realization of a straight corners (left) and inputs u3 and u2 for the realization of a straight corners (right). denote the command signals for z. 23] Medium. we use the architecture known as ‘‘mini-JEAN’’. the main advantage of the described technique is the optimization of the controller with respect to the actual characteristics of the engine. For this behavior. the reader is referred to the web version of the article. 32] Big and VB: [34.K. 6. B: [28. H.M. (For interpretation of the references to colour in the artwork. Table 2 illustrates the linguistic translation of the table obtained by on-line optimization for the z displacement (Table 1). 5. The optimization phase tends to stable weights (Fig. M: [19. NS: Negative Small. the building of the navigation controller is done entirely on-line by the optimization of a zero order Takagi–Sugeno fuzzy inference system (FIS) by a back-propagation-like algorithm. x or y directions respectively. Figs. 4 right): VW: [1. 559 The results of the simulation are reported in the Table 1 for z displacement.) . 4 left). In these circumstances the outputs linguistic labels could be interpreteted as follows (Fig. For the control of the XSF. The inputs u2 tend to zero after having carried out the desired orientation of the vehicle. It is also Fig.

(For interpretation of the references to colour in the artwork. the reader is referred to the web version of the article. Fig. Realization of a helical trajectory (left) and inputs u3 and u2 for the realization of a helical (right). XSF forces in the case of motor 4 failure at t ¼ 0:5 s (left) and t ¼ 4 s (right). In the case of the XSF. These figures show the effectiveness of the used controller.) . Controllers robustness The robustness of the developed controllers are evaluated regarding external disturbances and performance degradations in the case of actuator/sensor failures and wind influence. 8 which represents the realization of a helical trajectory (left) and its input (right) that our controller ensures the trajectory continuation. arc and round corners like connection (see Figs. Zemalache. 8. The 3D displacement is depicted with straight. The work produced by this force involves an additional energy consumption at the actuators levels which limits its maneuvering capacities in flight. This force can be expressed as follows: F i ¼ 12C x rAVi2 (29) Fig. 9. a resistance or a drag force is opposed to its movement in flight. Displacement errors of straight corners (left) and round corners (right).560 K. According to Fig. 7. 5.M. (For interpretation of the references to colour in the artwork.) shown in this figure that we can stabilize the system to make a following movement by the swivelling of the engine actuators 1 and 3. 5 and 6(left)). 7 shows displacement errors according to all the directions for the X4 Stationnary Flyer. It is noticed that the error thus tends to zero towards the desired positions. Fig. Maaref / Applied Soft Computing 9 (2009) 553–562 Fig. H. the reader is referred to the web version of the article.

[5] K. pp. R. [14] M. DC. J. (2003). Bouabdallah. AIAA. 301–314. Altug. pp. Modelling and simulation of a mini quad-rotor helicopter. [4] N. 2005. Real-time stabilization and tracking of a four rotor mini-rotorcraft. 600–605.4 N (left) and 2. June. (1995). pp.08 for the z displacement. 2693–2698. The STFIS controller exhibits chattering signal problems in the transition phase while the SFL controller presents static errors that varies proportionally with the drag force amplitude F dg . pp. We have presented and implemented an optimization technique allowing an on-line adjustment of the fuzzy controller parameters. Dzul. pp. A.Y. Slim. Streamlined rotors mini rotorcraft: trajectory generation and tracking. California. 10 present the simulation results in the case of a drag force of F dg ¼ 1:4 N and of F dg ¼ 2:1 N according to the x displacement.P. Quadrotor control using dual visual feedback. Bouabdallah. A comparison between the STFIS set rules and that deduced by human expertise. 10. Mahony. Indeed. 7–10. New Orleans. July. C.M. pp. in: IEEE International Conference on Robotics and Automation. vol. CA. References [1] E. 1. Obstacles avoidance and flying multi-drone are also envisaged thanks to the FIS faculties and its optimization capabilities. The objectives are to test the capability of the engine to fly with straight. 4294–4299. An evolutive neuro-fuzzy network. pp.K. Unmanned aerial vehicle speed estimation via nonlinear adaptive observers. Ostrowski. Benzemrane. in: Proceedings of the IEEE International Conference on Neural Networks. Spain. economical and safe since it is done on a mini-flying robots. Abichou. in: Fifth International Workshop on Robot Motion and Control. (2002). with its capacities to adapt to unknown situations by the means of its faculties of optimization. The same observations are found according to the two directions y and z. May. USA. vol. [9] C. September. [10] S. with its capacities of empirical knowledge modelling. Proceedings of the IEEE International Conference on Neural Networks. The descent gradient algorithm. Control of a quadrotor helicopter using visual feedback. Castillo. Vision based control of unmanned aerial vehicles with applications to an autonomous four rotor helicopter. Rohnert Park. Maneuvers for a Quad. September. in: Proceedings of the IEEE International Conference on Robotics and Automation. and the fuzzy logic. Abichou. Siegwart. Paris. arc. Poland. Castillo. R. Zemalache. Glorennec. Lozano. Maaref / Applied Soft Computing 9 (2009) 553–562 561 Fig. in: Proceedings of the 4e International Conference ‘‘Neural Networks and their Applications’’. Washington. (2005). R. Taipei. Bestaoui. Barcelona. Dzul. To maintain its equilibrium. the three remain actuators share the drone load compensation and which practically results in an equitable distribution of the developed forces (F 1 þ F 2 þ F 3 ¼ mg at steady state). Murrieri. P. [15] P. 2004. Orlando. The Eq. The surface characteristic A of the XSF drone is equal to A ¼ 0:031 m2 and it density is considered equal to r ¼ 1:22 kg/m3. Siegwart. 6. The STFIS and the SFL controllers behaves in the same way. . Noth. Lozano. J. Chow. USA.P. in: IFAC 15th World Congress on Automatic Control. R. Azouz. Taiwan. 2007. we studied a new configuration of flyer engine called XSF. It leads to very quick and efficient optimization technique.M. G. R. PID vs LQ control techniques applied to an indoor micro quadrotor. 72–77. Wind influence with a drag force of 1.Rotor Autonomous Helicopter. Dymaczewo. in: Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems. R. Zemalache. 141–148.M. R. This coefficient is equal to 0. 2259–2264. USA. (1991). Ostrowski. Smooth control of an X4 bidirectional rotors flying robots. 9 illustrate the simulation results in the case of actuator 4 failure after take-off at the instant t 1 ¼ 0:5 s and t 2 ¼ 4 s in the z direction. Mahony. [13] P. IEEE Transactions on Control Systems Technology 12 (July (4)) (2004) 510–516. [3] E. (2007). shows the advantages and the disadvantages of these two techniques. pp. A. [8] Y. Stabilization of a mini-rotorcraft having four rotors.P. Regularization and complexity control in feedforward neural networks. Barcelona. An analysis of weight decay as a methodology of reducing three-layer feedforward artificial neural network for classification problems. Fig. Future works will essentially investigate the real time implementation of the STFIS and the based-model control techniques.5 for the x and y directions and 0. 1. in: American Control Conference. shows the validity of the proposed technique. in: Proceeding of the IEEE International Conference on Robotics and Automation. H. 2003. Santosuosso. April. [7] L. (For interpretation of the references to colour in the artwork. Hamel. [16] T. Bishop. Y. Damm. 2005. [12] P. A. Dynamic modelling and configuration stabilization for an X4-flyer. the reader is referred to the web version of the article. we have obtained an on-line optimized Takagi–Sugeno type FIS of zero order. 2002. Lozano. Long Beach. Beji. Nıˆmes. pp. in: ICNN’94. 11–13.) where F i [N] is the drag force following the i axis. A. Altug. International Journal of Automation and Systems 3 (1 (March)) (2005) 87–99. Sendai. [6] L. J. Fig. Altug. Quadrotor. Japan. [2] E.L. K. in: Proceeding of the IEEE International Conference on Robotics and Automation. Bestaoui. G. (1995). Conclusion In this paper. Pennsylvania. rounded intersections and complex trajectories (helical). An analysis of the STFIS (which not require the good knowledge of the model) and the SFL (requires the well knowledge of the system model and parameters) controllers and their robustness regarding disturbances. A [m2] is the coss-sectional area perpendicular to the force flow and r [kg/m3] is the body density. A. ASME Design Theory and Methodology Coferences. (29) induced a drag coefficient C x which is a dimensionless quantity that describes a characteristic amount of aerodynamic drag depending on the XSF structure and which is experimently determined by windtunnel tests. Spain. are combined to control a new configuration of flyer engine. in: Proceeding of DTM2005. This method is simple. Beji. R. We have considered in this work the stabilizing/ tracking control problem for the three decoupled displacements of a XSF. (2004). V i [m/s] is the drone velocity. Ostrowski. [11] S. Faculties of the University of Pennsylvania.1 N (right) for the x direction. New York City. Taylor. Design and control of an indoor micro quadrotor.

vol.M. 2005. Derivation of fuzzy control rules from human operator’s control actions. Rajnarayan. [26] H.L. Zemalache. Robotics and Autonomous Systems 38 (2002) 1–18. Self-turnable fuzzy inference system: a comparative study for a drone.562 K. Demblewski. Wakami.L. pp. reinforcement learning. [32] H. H. Mexico.P. Lai. 2005. 21–24. L. [31] V. St. Zemalache. McGilvray. [45] K. pp. E. Sensor-based navigation of a mobile robot in an indoor environment. Maaref. pp.M. Vancouver. Automatica 28 (4) (1992) 665–679. Maaref. S. . G. Approche connexionniste de la commande de processus. 2005.N. Journal of Electrical Engineering 57 (1) (2006) 20–27. Cancun. 1. Nonlinear control design for slightly non-minimum phase systems: application to V/STOL aircraft. Nomura. Andrews. Genova. 151-156. C. 2007. in: IEEE/RSJ International Conference on Intelligent Robots and Systems. Waslander. Daachi. Ichihashi. [36] T. A. J. [38] G. 70–74. Benallegue. ANFIS: adaptive-network-based fuzzy inference system. Bruxellese. Internal world models and supervised learning.J. AIAA2005-7077. Beji. [34] J. pp. Pure 2D Visual Servo control for a class of under-actuated dynamic systems. IEEE Transactions System. pp. L. 2004. Mokhtari. Maaref. Hauser. N. MO. S. C. 160–230. Hoffmann. A. Navigation. Jang. International Conference on Fuzzy System. N. L. Zemalache. (2006). Algeria. 2001. Control of an under-actuated system: application to a four rotors rotorcraft. (2004). Thse de doctorat.J. Salt Lake City. New concepts and perspectives on micro-rotorcraft and small autonomous rotarywing vehicles. J. A small semi-autonomous rotary-wing unmanned air vehicle (UAV). Iterative fuzzy modelling and a hierarchical network. Estimation and control of a quadrotor vehicle using monocular vision and moire´ patterns. Man and Cybernet 23 (1993) 665–685. Valentiy ve. (1991). Mistler. A self-tuning method of fuzzy control by descent method. R. 21–24. Progressive Optimization of a Fuzzy Inference System. New Orleans. [19] S.S. 295–300. D. 1991. October. Horn. UT. Benallegue. pp. August. Tayebi. [33] P. in: Proceedings of the 23rd Digital Avionics System Conference. J. pp.A. Saerens. April. K. Roberts. pp. Hynes. International Conference on Advances in Vehicle Control and Safety. Dostal. Tomlin.M. [29] A. Jang. R. [20] G. Attitude stabilization of a VTOL quadrotor aircraft. USA. [43] K. J. Beji. Maaref. Mahony. IFSANAFIPS’2001. Rumelhart. C. July. [22] H. Hoffmann. June 29–July 03. Hayashi. USA.W. in: Proceedings of the IFAC Symposium On Fuzzy Information. Beji. M. 49–52. Renders. Yang. Meyer. [41] C. pp. Klem. C. J. 1991.G. Waslander. 665–679. Pound. 2002. 2359–2366. in: Proceedings of the 1st International in Computer System and Information Technology Conference. S. in: Proceedings of the IEEE International Conference on Intelligent Robots and Systems. Tomlin. July. [35] M. pp. Thse de Doctorat de l’Universit Libre de Bruxellese. M. M’sirdi.M. Tournier. and Control Conference and Exhibit.I. Barret. H. [30] A. Wang. with quadrotor applications.P. 2006. [37] A. 1994. Navigation. 2005. Barret.C. 1. 155–158. [42] L.F. L. (2004). Bordeaux and Paris. in: Proceedings of the eighth International Workshop on Machine Learning. (1991). Mederreg. pp. H. J. [23] J. Auckland. 145–150. [18] J. in: Proceedings of the Australasian Conference on Robotics and Automation. Keystone. Johnson. Universit libre de Bruxellese. Italy. in: IEEE ICSS International Conference On Systems and Signals. 2229–2235. vol. J. Design of a four rotor aerial robot. Maaref.M. Exact linearization and noninteracting control of a 4 rotors helicopter via dynamic feedback. in: International Fuzzy Systems Association (IFSA). Wu. Bruxellese. J. 28–31. 2004. November. S. Hanford. Brussels. China. Back-propagation fuzzy system as nonlinear dynamic system identifiers. [39] L. The stanford testbed of autonomous rotorcraft for multi-agent control (STARMAC).L. [25] H.X. in: AIAA Guidance. N. August. CO. Multi-agent X4-flyer testbed control design: integral sliding mode vs. [24] M.M. pp.K. in: Proceedings of the 4th IFSA Congress. Benallegue. [44] K.S. Sastry. CO. vols. Brussels. Sugeno.D. Takagi. in: Proceedings of the 20th AIAA Applied Aerodynamics Conference. 1409–1416. and Control Conference and Exhibit. LA. 2001. Distributed cooperative search using information— theoretic costs for particle filters. L. Time-optimal control of a hovering quad-rotor helicopter. in: Proceedings of the 5th IEEE International Conference on Robotics and Biomimetics. J. Aiken. F. in: AVCS’04. Long. [28] L. Backstepping control technique: application to an under-actuated X4-flyer. September. 18–21. Zemalache. vols. Mahony. D. in: Proceedings of the 2002 IEEE International Conference on Robotics and Automation.L. Beji. Nonlinear backstepping control with observer design for a 4 rotors helicopter. D. Maaref / Applied Soft Computing 9 (2009) 553–562 [17] T. in: Proceedings of the 4th IFSA Congress. in: Proceedings of the IEEE. Zemalache. Hamel. M’Sirdi. Robust feedback linearization and GH1 controller for a quadrotor unmanned aerial vehicle. R. Louis. P. Young. 2002. 468–473. G. Hong Kong. Maaref. C. in: IEEE 10th IEEE International Workshop on Robot–Human Interactive Communication. Keystone. San Diego. 55–60. [21] G. June 18–21. AIAA2006-6711. Howz. Journal of the Integrated Computer-Aided Engineering (ICAE) 14 (January 2) (2007) 107–119.R. Dynamic feedback controller of Euler angles and wind parameters estimation for a quadrotor unmanned aerial vehicle. Knowledge Representation and Decision Analysis. in: AIAA Guidance. H. motion planning and control. B. Waslander. A. Jordan. Hoffmann. Diaz. Jang. [27] H. IEEE Transactions on Control Systems Technology 14 (May (3)) (2006). L. [40] S. 1983. Rendel.K. Mokhtari. I. (1991). Mtaphore biologiques appliques a´ la commande de processus. Two inertial models of X4-flyers dynamics.M.