Professional Documents
Culture Documents
Correct-by-Construction Navigation Functions With Application To Sensor Based Robot Navigation
Correct-by-Construction Navigation Functions With Application To Sensor Based Robot Navigation
Abstract— This paper brings together the concepts of naviga- • A correct-by-construction Navigation Function
tion transformation and harmonic functions to form navigation • On-the-fly addition of obstacles in the workspace
functions that are correct-by-construction in the sense that • A relaxed version of the Morse Property of the Navi-
no tuning is required. The form of the navigation function is
explicitly related to the number of obstacles in the environment. gation Functions that provides equivalent performance
characteristics
arXiv:2103.04445v1 [cs.RO] 7 Mar 2021
its Taylor expansion converges on any open set of its domain. where S is the boundary of a ball centered at θd with radius r
Let E be an open set containing a critical point hc of φk . and σ the normalized perimeter measure. Since hd is isolated,
Assume that the critical point is not isolated. This implies φk on S cannot be identically θd . In particular for the mean
that there is either (a) an open neighborhood Z around hc value on S to be θd , φk needs to take both higher and lower
where ∀hz ∈ Z : φk (hz ) = φk (hc ) or (b) there is a 1-D manifold values than θd which implies that hd is a saddle point.
L , where ∀hL ∈ L : φk (hL ) = φk (hc ). The next Lemma establishes the nature of the basins of
Let us now examine case (a) where ”there is an open attraction of the critical points:
neighborhood Z around hc where ∀hz ∈ Z : φk (hz ) = φk (hc ) ”. Lemma 4: The basin of attraction of the saddle points of
Then the Taylor expansion of φk around hc will be: φk is a set of measure zero.
( ) Proof: Since Φ(·) is a diffeomorphism, it is sufficient
1 n
∞
n to prove this for φk . Assume a basin of attraction Ω of the
φk (h) = φk (hc ) + ∑ ∑ m Dn,m |hc hn−m m
cx hcy
critical point hc ∈ C (φ ) such that µ(Ω) > 0 and hc ∈ ∂ Ω,
n=1 n! m=0
n
where µ(·) is the Lebesque measure of a set. Then Ω is a
where Dn,m = ∂ xn−m ∂ φk
∂ ym
and h − hc = [hcx hcy ]T . Since positively invariant set with respect to the negated gradient
φk (hz ) = φk (hc ), ∀hz ∈ Z, this necessarily implies that ∀n > flows of φk , which will be pointing inward along ∂ Ω, i.e.
0, m ≥ 0 : Dn,m = 0. Since φk is analytic, for any open ∇φk (hω ) · n̂ ≥ 0, ∀hω ∈ ∂ Ω, where n̂ the outward pointing
connected component G of the domain of φk containing Z normal to set’s boundary. Take Ωn ⊆ Ω to be the largest
as a subset, it will hold φk (hg ) = φk (hc ), ∀hg ∈ G, which subset of Ω that contains hc and for which ∇φk (hω ) · n̂ > 0,
according to the Principle of Permanence implies that φk ∀hω ∈ ∂ Ωn \ hc . Such a set is guaranteed to exist as long as
will be identically constant. However this is not true and by φk is not flat in Ω. Hence
contradiction case (a) is not possible. I
Let us now proceed to examining case (b) where ”there is a ∇φk · n̂ dl > 0 (11)
∂ Ωn
1-D manifold L , where ∀hL ∈ L : φk (hL ) = φk (hc )”. We can
Using the 2-dimensional Divergence Theorem we have that:
now construct an analytic function λ : R≥0 → L with λ (0) = I Z
hc and such that λ (`) = h` where ` is the length of the curve ∇φk · n̂ dl = ∇ · ∇φk dA (12)
from hc to h` . Define the function φ L (·) , (φk ◦ λ )(·). Since ∂ Ωn Ωn
φ L : R≥0 → R is a composition of analytic functions, it will where dA the area differential. Since φk is a Harmonic func-
also be analytic. However by construction φ L (`) = φk (hc ), tion,
H
then by definition ∇ · ∇φk = ∆φk = 0 and consequently
and this will be true for its whole domain ` ∈ R≥0 according ∂ Ωn ∇φ k · n̂ dl = 0. This is in contradiction with eq. (11) and
to the Principle of Permanence. This implies that L will since φk in Ω is non-flat, Ω reduces to a set of measure zero.
have an infinite length. Then L will have infinitely large
Theorem 1: Given a smooth navigation transformation V. NAVIGATION F UNCTION C ONTROLLER D ESIGN
Φ(·), the function ϕk (x) = σ ◦ φk ◦ Φ(x) is a navigation A. Setup
function on W as per Definition 4 for k > M.
In the previous section we have shown that ϕk where
Proof: Let us examine the properties of Definition 4.
k ≥ M + 1 is a Navigation Function where M is the number
Property 1: The function ϕk (·) is smooth since it is a
of workspace obstacles. As the robot moves in the environ-
composition of smooth functions.
ment, the number M of obstacles encountered is increasing.
Property 2: From Lemma 1 we have that φk is free of
Let d be the detection radius of the robot. Any obstacle
local minima. From (5), using the chain rule, we have that
entering this radius is registered as a new obstacle and the
∇ϕk = ( dσ T
dφ ◦ φk ◦ Φ)JΦ (∇φ ◦ Φ), where JΦ is the Jacobian number of obstacles M increases when fist detected. Let
of Φ. Since Φ is a diffeomorphism, JΦ is non-singular and
T = {T1 , . . . , TN } be the time instants where new obstacles
since σ is strictly increasing, away from the destination xd
are discovered during the navigation of the robot and assume
the set of critical points in W and the corresponding point
that N is the total number of obstacles that are discovered
world match, C (φk ) = Φ(C (ϕk ) − {xd }). Also at the critical
during a navigation instance. Let n(t) denote the number of
points away from xd we have that: ∇2 ϕk = ( dσ
dφ ◦ φk ◦
Cφ
k
obstacles that were discovered up to the time instant t and
Φ)JΦT (∇2 φ ◦ Φ)JΦ . Hence, using the same arguments as in define:
[6], away from xd we have index(ϕ)|Cϕ = index(φ )|Cφ . Θ(x,t) , ϕk(t) (x) (16)
k k
Since φk has no local minima in P, the same will be true where k(t) > n(t) an integer1 . Eq. (16), is a continuous
for ϕ in W . Now let us examine xd . From Eq. (5) and using (smooth) function in the state but discontinuous in time,
the notation h = Φ(x) we have that: representing in every time interval Tn(t) ≤ t < Tn(t)+1 the
Navigation Function that is applied to the discovered en-
kh − Pd k2
ϕk (h) = vironment at time t.
M 2
kh − Pd k2 + ∏ kh − Pi k k To ensure that the Navigation Transformation Φ is well
i=1 defined at any time instant, we will have the following:
2(h−Pd )D−kh−Pd k ∇D 2 Assumption 1: Sensed obstacles’ shapes and orientations
Hence ∇ϕk (h) = D2
where D = kh − Pd k2 + are known.
M 2
∏ kh − Pi k k . The Hessian of ϕk at xd is thus In addition we need to have an obstacle ”memory” in our
i=1 system as follows:
2 Assumption 2: Obstacles become part of the known
Hϕk (xd ) = I. (13)
M 4 workspace after being sensed.
∏ kPd − Pi k k
Considering the definition of the symmetric sensing sector
i=1
Sr,θ , it can be shown that the minimum distance that a robot
The destination xd is thus the unique minimum of ϕk in W . can approach to an obstacle before detecting it, is given by:
Property 3: By Proposition 2 the critical points of φk are
isolated. As discussed in the previous, C (φk ) = Φ(C (ϕk ) − θ ρmin
min r sin 2 , θ , θ < π
{xd }). Hence ϕk has no additional critical points except xd , dmin (θ ) = cos 2 (17)
r, π ≤ θ ≤ 2π
which was shown to be a non-degenerate global minimum.
Hence all critical points of ϕk are isolated. By Lemma 3 where ρmin is the minimum radius of curvature of the
the isolated critical points except xd are saddle points and workspace obstacles.
by Lemma 4 the basin of attraction of these critical points The following assumption is about knowledge of the
has zero measure. Since Φ(·) is a diffeomorphism, the above starting neighborhood:
properties will also hold for ϕk (x). Assumption 3: For a symmetric sensing sector Sr,θ , any
Property 4: By construction lim φk (h)=∞ for i = 1 . . . M. obstacles that are within a radius of dmin (θ ) of the initial
h→Pi
Moreover, we have the requirement that lim kΦ(x)k = +∞ robot configuration, are assumed known.
x→O0
which in combination with Lemma 2 implies that B. Kinematic Controller Design
Let us now consider the first-order kinematic system
lim φk ◦ Φ(x) = +∞. (14) model of Eq. (1). Our motivation in setting up the control law
x→∂ W
for this system, is to create a vector field that provides kinetic
From Eq. (4) we have that energy to the system that matches (or is proportional to)
the level of the Navigation Function. We have the following
lim σ (x) = 1. (15) result:
x→∞
Proposition 3: System (1) with a symmetric sensing sec-
Combining eqs. (14) and (15)) we obtain tor Sr,θ under the control law:
√
lim ϕ = 1 u = −K 2Θ · ∇Θ (18)
x→∂ W
d
zero. Hence by LaSalle’s Invariance Principle the system λ = λc = 2 2µm ∏ kPd − Pi k k (21)
i=1
will converge to the largest invariant set that includes the
destination configuration and the saddle points that are the provides critical damping at the neighborhood of the desti-
ω limit set of a set of measure zero of initial conditions. nation configuration.
a.e. Proof: From Eq. (13), the Navigation Function µΘ will
Hence V̇ < 0 and the control law of Eq. (18) is globally
asymptotically stable, almost everywhere. behave locally at the destination neighborhood as an elastic
Case 2: Local sensing i.e. r is finite and 0 < θ ≤ 2π: potential, i.e. µΘ ≈ 12 ksp xT x, with spring constant:
In this case, the robot senses an obstacle’s existence before M −2
the minimum distance to the obstacle becomes dmin (θ ). ksp = 2µ ∏ kPd − Pi k k
Let t = Tν be the time instant that the robot discovered the i=1
ν’th obstacle. At this time instant the control law (18) will Since λ acts as the damping coefficient for system (2) under
have a discontinuity in time. This discontinuity will be an control law (20), the damping ratio of the second order
isolated event due to Assumption 2. Hence the trajectories system will be:
of System (1) under control law (18), will flow along λ
ζ= p .
Carathéodory solutions, and following the same Lyapunov 2 mksp
analysis as the previous case, we get: Setting ζ = 1 for critical damping, we get the result in Eq.
n √ o (21)
ess sup −K 2Θ k∇Θ(x,t)k ≤ 0 We can now state the following result in the case of local
t
a.e. sensing:
which implies V̇ < 0 and the control law of Eq. (18) is Proposition 5: System (2), with a symmetric sensing sec-
globally asymptotically stable, almost everywhere. tor Sr,θ under the control law:
C. Dynamic Controller Design f = −µ∇Θ − λ (t)ẋ (22)
Now let us consider the case of the dynamical System (2).
with
Let
1
2√2µm kP − P k −1
n(t)
V (t) = µΘ + mẋT ẋ (19)
2 ∏ d i k , V <µ
λ (t) = i=1 (23)
be the total energy of the system. In this case our system
m
ẋ(Tn(t) ) , V ≥µ
dmin(θ )
will be converting potential energy to kinetic energy and
vice-versa while dissipating energy through an appropriate and µ > 0, r > 0 and 0 < θ ≤ 2π ensures collision avoid-
dissipation term. We have the following result: ance everywhere, critical damping at the neighborhood of
Proposition 4: System (2) with global sensing, under the the destination configuration and, almost everywhere global
control law: asymptotic stability.
f = −µ∇Θ − λ ẋ (20) Proof: System (2) under control law (22) can be
considered as a mechanical system with dissipation under
with µ, λ positive gains, is globally asymptotically stable
a conservative force field. This implies that if the system
almost everywhere.
starts from rest, then collisions are not possible as long as
Proof:
the total energy of the system is bounded below µ. This is a
Using the same reasoning as in the Case 1 of the proof
direct result of the admissibility Property of the Navigation
of Proposition 3 and choosing V as a Lyapunov function
Function.
candidate we have that:
However, it is possible that the jump in the potential level
V̇ = µ∇T Θ · ẋ + ẋT · (−µ∇Θ − λ ẋ) = −λ kẋk2 ≤ 0 of Θ, right after a new obstacle is encountered, can cause the
total energy to go above µ. In this case we need to ensure In the first simulation study, the kinematic controller (18)
that, for the worst case scenario, there is sufficient dissipation was applied to the system (1). The robot started at its
in the system to absorb the energy added due to introduction initial position in a ”seemingly” empty workspace. As the
of a new obstacle. robot moved, new obstacles were discovered through the
Now assume that the robot is heading towards the newly symmetric sensing sector as shown in Fig. 1, starting from
discovered obstacle. To simplify our analysis let us not take O1 up to O6 . As can be seen the proposed controller per-
into account the repulsive force generated by the navigation formed successfully, maintaining the robot in the workspace,
function and assume that the robot is moving towards the handling the occurrence of new obstacles, avoiding collisions
obstacle with only the dissipation force slowing it down. and stabilizing at the destination configuration.
Then the differential equation describing the motion of the
robot is given by mẍ + λ ẋ = 0 where x is the distance across
the straight line connecting the robot position x(Tn(t) ) when Start
the obstacle was detected and the closest obstacle point. The Workspace
solution to the differential equation is boundary
m O O3 O2
λ 4
x(t) = kẋ(Tn(t) )k 1 − exp− m t (24)
λ O1
and its time derivative is
λ
ẋ(t) = kẋ(Tn(t) )k exp− m t (25) O5
ACKNOWLEDGMENT
The first author would like to acknowledge the contri-
bution of European Union’s Horizon 2020 research and
innovation program under grant agreements 824990 (RIMA),
767642 (L4MS) and by the Cyprus Research and Innova-
tion Foundation grants EXCELLENCE/1216/0365 (HOD-
ICCCS) and EXCELLENCE/1216/0296 (RETuNE).
R EFERENCES
[1] Omur Arslan and Daniel E Koditschek. Sensor-based reactive naviga-
tion in unknown convex sphere worlds. The International Journal of
Robotics Research, 38(2-3):196–223, 2019.
[2] M. Audin and M. Damian. Morse Theory and Floer Homology.
Springer-Verlag, London, 2014.
[3] S. Axler, P. Bourdon, and W. Ramey. Harmonic Function Theory.
Springer-Verlag New York, Inc., 2nd edition, 2001.
[4] N. Constantinou and S. G. Loizou. Robot navigation on star worlds
using a single-step navigation transformation. 59th IEEE Conference
on Decision and Control (CDC), pages 1537–1542, 2020.