10 1 1 92 9625

Geom.Ctrl.Mech.
Sys-ME225FB-W06-p1
Geom.Ctrl.Mech.Sys-ME225FB-W06-p2
Introduction Geometric Control of Mechanical Systems

UCSB Mechanical Engineering Winter 2006, ME 225 FB: Special Topic Course Focus: Geometric Control of Lagrangian Systems Some sample systems
F1 r
F2
Francesco Bullo
Andrew D. Lewis

Mechanical Engineering Mathematics & Statistics University of California at Santa Barbara Queens University Kingston bullo@engineering.ucsb.edu, http://motion.mee.ucsb.edu

Thanks to: Thanks to:
Jorge Corts, Kevin Lynch, Sonia Mart e nez, Milo Zefran s

National Science Foundation
Sample problems (concretely) Start from rest. Sample problems (vaguely) Modeling: Is it possible to model the four systems in a unied way, that allows for the development of eective analysis and design techniques? Analysis: Some of the usual things in control theory: stability, controllability, perturbation methods. Design: Again, some of the usual things: motion planning, stabilization, trajectory tracking. (i) Describe the set of reachable states. (a) Does it have a nonempty interior? (b) If so, is the original state contained in the interior? (ii) Describe the set of reachable positions. (iii) Provide an algorithm to steer from one position at rest to another position at rest. (iv) Provide a closed-loop algorithm for stabilizing a specied conguration at rest. (v) Repeat with thrust direction xed.

1.3
Motion planning
1
1.1 Scientic Interests
Broad motivations
Example systems (i) dextrous manipulation via minimalist robots (ii) real-time trajectory/path planning for autonomous vehicles (iii) locomotion systems (walking, swimming, diving, etc)
(i) success in linear control theory is unlikely to be repeated for nonlinear systems. In particular, nonlinear system design. no hope for general theory mechanical systems as examples of control systems (ii) control relevance of tools from geometric mechanics (iii) geometric control past feedback linearization
Application contexts 1.2 Industrial Trends new concepts in design implementation on-line focus on actuators and algorithms (i) guidance and control of physical systems (ii) prototyping and verication (iii) graphical animation and movie generation (iv) analysis of animal and human locomotion and prosthesis design in biomechanics
autonomous vehicles recongurable, reactive sensing & computation cheap
The literature, historically Godbillon [1969], Abraham and Marsden [1978], Arnold [1978]: Geometrization of mechanics in the 1960s. Nijmeijer and van der Schaft [1990], Jurdjevic [1997], Agrachev and Sachkov [2004]: Geometrization of control theory in the 1970s, 80s, and 90s by Agrachev, Brockett, Hermes, Krener, Sussmann, and many others. Brockett [1977]: Lagrangian and Hamiltonian formalisms, controllability, passivity, some good examples. Crouch [1981]: Geometric structures in control systems. van der Schaft [1981/82, 1982, 1983, 1985, 1986]: A fully-developed Hamiltonian foray: modeling, controllability, stabilization. Takegaki and Arimoto [1981]: Potential-shaping for stabilization. Bonnard [1984]: Lie groups and controllability. The literature, historically (contd) Bloch and Crouch [1992]: Ane connections in control theory, controllability. Koiller [1992], Bloch et al. [1996], Bates and Sniatycki [1993], van der Schaft and Maschke [1994]: Geometrization of systems with constraints. Bloch et al. [1992]: Controllability for systems with constraints. Baillieul [1993]: Vibrational stabilization. Ortega et al. [1998], Arimoto [1996]: Texts on stabilization using passivity methods. Bloch et al. [2000, 2001], Ortega et al. [2002]: Energy shaping. Bloch [2003]: Text on mechanics and control.
What we will try to do this today Present a unied methodology for modeling, analysis, and design for mechanical control systems. The methodology is dierential geometric, generally speaking, and ane dierential geometric, more specically speaking. Follows: Geometric Control of Mechanical Systems: Modeling, Analysis, and Design for Simple Mechanical Control Systems Francesco Bullo and Andrew D. Lewis SpringerVerlag, 2004, ISBN 0-387-22195-6, http://penelope.mast.queensu.ca/smcs Warning! This lecture series will be much less precise than the book. We do not claim that the methodology presented is better than alternative approaches. Todays topics: Lecture #1: Geometric Modeling Lecture #2: Controllabilty Lewis and Murray [1997] Lecture #3: Kinematic Reduction and Motion Planning Bullo and Lynch [2001], Bullo and Lewis [2003] Additional Lecture: Perturbation methods and oscillatory stabilization Bullo [2002, 2001], Mart nez et al. [2003]
Outline
Geometric modeling of mechanical systems

Dierential geometry essential: Advantages (i) Prevents articial reliance on specic coordinate systems. (ii) Identies key elements of system model. (iii) Suggests methods of analysis and design. Disadvantages (i) Need to know dierential geometry. Manifolds
M a Ua Rn
Manifold M, covered with charts {(Ua , a )}aA satisfying overlap condition. Around any point x M a chart (U, ) provides coordinates (x1 , . . . , xn ). Continuity and dierentiability are checked in coordinates as usual.
Ub
b ab
Rn
Tangent bundles Manifolds (contd) Manifolds we will use this week: (i) Euclidean space: Rn . (ii) n-dimensional sphere: Sn = {x Rn+1 | x (iii) m n matrices: R
mn Rn+1
0 U x [1 ]x = [2 ]x 1 M
= 1}.
2 Rn
Formalize the idea of velocity.
(iv) General linear group: GL(n; R) = { A Rnn | det A = 0}. (v) Special orthogonal group: SO(n) = {R GL(n; R) | RRT = I n , det R = 1}. (vi) Special Euclidean group: SE(n) = SO(n) R . The manifolds Sn , GL(n; R), and SO(n) are examples of submanifolds, meaning (roughly) that they are manifolds contained in another manifold, and acquiring their manifold structure from the larger manifold (think surface).
n
Given a curve t (t) represented in coordinates by t (x1 (t), . . . , xn (t)), its velocity is t (x1 (t), . . . , xn (t)). Tangent vectors are equivalence classes of curves. The tangent space at x M: Tx M = {tangent vector at x}. The tangent bundle of M: TM = xM Tx M. The tangent bundle is a manifold with natural coordinates denoted by ((x1 , . . . , xn ), (v 1 , . . . , v n )).
Vector elds Assign to each point x M an element of Tx M.

M Rn
Flows Vector eld X and chart (U, ) . . . xn (t) = X n (x1 (t), . . . , xn (t)). Solution of o.d.e. curve t (t) satisfying (t) = X((t)). o.d.e.:
x1 (t) = X 1 (x1 (t), . . . , xn (t))
Coordinates (x1 , . . . , xn )
vector elds { x1 , . . . , xn } on chart domain. X i xi
Such curves are integral curves of X. Flow of X: (t, x) X (x) where t X (x) is the integral curve of X t t through x.
Any vector eld X is given in coordinates by X = summation convention).
(note use of
Lie bracket Flows do not generally commute. i.e., given X and Y , it is not generally true that X Y = Y X . t s s t The Lie bracket of X and Y : d [X, Y ](x) = dt
Y X Y t Xt (x). t t
Vector elds as dierential operators Vector eld X and function f : M R X: d L X f (x) = dt Lie derivative of f with respect to f (X (x)). t
t=0
f In coordinates: L X f = X i xi (directional derivative).
t=0
Measures the manner in which ows do not commute. Mechanical exhibition of the Lie bracket
One can show that L X L Y f L Y L X f = L [X,Y ] f [X, Y ] = Y i j X i j X Y . xj xj xi
Conguration manifold
b2
b1 Obody
Conguration manifold (contd) Most systems are not free, but consist of bodies that are interconnected.
s3
r b3 s2
Ospatial
Single rigid body:
s1
positions of body
(Obody Ospatial ) R3 b1 b2 b3 SO(3).
Denition 1. An interconnected mechanical system is a collection B1 , . . . , Bk of rigid bodies restricted to move on a submanifold Q of Qfree . The manifold Q is the conguration manifold. Coordinates for Q are denoted by (q 1 , . . . , q n ). Often called generalized coordinates. For j {1, . . . , k}, j : Q SO(3) R3 gives conguration of jth body. This is the forward kinematic map.
Q = SO(3) R3 for a single rigid body. For k rigid bodies, Qfree = (SO(3) R3 ) (SO(3) R3 )
k copies
This is a free mechanical system.
Conguration manifold (contd) Example 2. Planar rigid body: Q = SO(2) R2 S1 R2 .

b2 b1 (x, y) Obody s2
Conguration manifold (contd) Example 3. Two-link manipulator: Q = SO(2) SO(2) S S .

s2
b2,2
b2,1 2 b1,2 b1,1 1 s1
Coordinates (1 , 2 ). Coordinates (, x, y). cos sin 0 sin cos 0 0

Ospatial s1
1 (, x, y) =
=R1 SO(3)
0, (x, y, 0) . 1
=r 1 R3
1 (1 , 2 ) = (R1 , r 1 ) and 2 (1 , 2 ) = (R2 , r 2 ), where cos 1 sin 1 sin 1 R1 = cos 1 0 0 r 1 = r1 R 1 s 1 , r2 =
1 R1 s1
0 , 1
+ r2 R 2 s 1 .
cos 2 sin 2 R2 = 0
sin 2 cos 2 0
0 , 1
Conguration manifold (contd) Example 4. Rolling disk: Q = R2 S1 S1 .

s3 b3 s2 s1 (x, y) b2 b1
Velocity Rigid body B undergoing motion t (R(t), r(t)): (ii) Spatial angular velocity: t (t) (i) Translational velocity: t r(t); R(t)R1 (t); R1 (t)R(t).
(iii) Body angular velocity: t (t)
Coordinates (x, y, , ).
1 (x, y, , ) =
cos cos cos sin sin
sin cos sin sin cos

=R1 SO(3)
cos , (x, y, ) . 0
=r 1 R3
sin
Both (t) and (t) lie in so(3) dene (t), (t) R3 by the rule 0 a3 a2 a3 (a1 , a2 , a3 ). 0 a1 a2 a1 0
Inertia tensor Rigid body B with mass distribution . Mass: (B) =

B
Kinetic energy Rigid body B undergoing motion t (R(t), r(t)). Assume Obody is at the center of mass (xc = 0). Kinetic energy:
3
d.
B
Centre of mass: xc =
x d.
3
Inertia tensor about xc : Ic : R R dened by Ic (v) =

B
1 r(t) + R(t)x 2 B Proposition 5. KE(t) = KEtrans (t) + KErot (t) where KE(t) = KEtrans (t) = 1 (B) r(t) 2
2 R3 ,
2 R3
(x xc ) (v (x xc )) d.
KErot =
1 2
Ic ((t)), (t)
R3
Kinetic energy (contd) Interconnected mechanical system with conguration manifold Q. vq TQ. t (t) Q a motion for which (0) = vq . jth body undergoes motion t j (t) = (Rj (t), r j (t)). Dene j (t) = R1 (t)Rj (t). j Dene KEj (vq ) = 1 j (Bj ) r j (t) 2
2 R3 1 2
Symmetric bilinear maps Need a little algebra to describe KE. Let V be a R-vector space. 2 (V) is the set of maps B : V V R such that (i) B is bilinear and (ii) B(v1 , v2 ) = B(v2 , v1 ). Basis {e1 , . . . , en } for V: Bij = B(ei , ej ), i, j {1, . . . , n}, are components of B.
R3 t=0
Ij,c (j (t)), j (t)
[B] is the matrix representative of B. An inner product on V is an element G of 2 (V) with the property that G(v, v) 0 and G(v, v) = 0 if and only if v = 0.
This denes a function KEj : TQ R which gives the kinetic energy of the jth body. The kinetic energy is the function KE(vq ) =
k j=1
KEj (vq ).
Example 6. V = Rn , GRn the standard inner product, {e1 , . . . , en } the standard basis: (GRn )ij = ij .
Kinetic energy metric (contd) Example 8. Planar rigid body: 0 I1,c = 0 , 0 0 J
Kinetic energy metric Proposition 7. There exists an assignment q G(q) of an inner product on Tq Q 1 with the property that KE(vq ) = 2 G(q)(vq , vq ). G is the kinetic energy metric and is an example of a Riemannian metric. G is a crucial element in any geometric model of a mechanical system.
1 (t) = (R1 (t)R1 ) = (0, 0, ), 1
1 KE = 2 m(x2 + y 2 ) + 1 J 2 , 2 J 0 0 [G] = 0 m 0 . 0 0 m
Kinetic energy metric (contd) Example 9. Two-link manipulator: 0 I1,c = 0 , 0 0 J1 Kinetic energy metric (contd) = 0 0 0 Example 10. Rolling disk: Jspin 0 0 I1,c = 0 Jspin 0 , 0 0 Jroll
I2,c
J2
1 (t) = (R1 (t)R1 ) = (0, 0, 1 ), 1
0 ,
1 (t) = (R1 (t)R1 ) = ( sin , cos , ), 1
2 (t) = (R1 (t)R2 ) = (0, 0, 2 ), 2

1 2 2 KE = 8 (m1 + 4m2 ) 2 1 + 1 m2 2 2 1 2 8 1 2 2 cos(1 2 )1 2 + 2 J1 1 + 1 J2 2 , 2 1 J1 + 1 (m1 + 4m2 ) 2 2 m2 1 2 cos(1 2 ) 1 4 . [G] = 1 m2 1 2 cos(1 2 ) J2 + 1 m2 2 2 2 4 1 + 2 m2
1 2
1 KE = 1 m(x2 + y 2 ) + 2 Jspin 2 + 1 Jroll 2 , 2 2 m 0 0 0 0 m 0 0 [G] = . 0 0 Jspin 0 0 0 0 Jroll
Kinetic energy metric (contd) This whole procedure can be automated in a symbolic manipulation language. Snakeboard example:

Euler-Lagrange equations Free mechanical system with conguration manifold Q and kinetic energy metric G. Question: What are the governing equations? Answer: The EulerLagrange equations.
1 Dene the Lagrangian L(vq ) = 2 G(vq , vq ).
Choose local coordinates ((q 1 , . . . , q n ), (v 1 , . . . , v n )) for TQ. The EulerLagrange equations are d L L i = 0, i dt v q

i {1, . . . , n}.
Here Q = R S S S with coordinates (x, y, , , ).
The EulerLagrange equations are rst-order necessary conditions for the solution of a certain variational problem.
EulerLagrange equations
1 Let us expand the EulerLagrange equations for L = 2 Gij (q)q i q j :
Ane connections 1 Glm l m Gkl qq m q 2 q k Denition 11. An ane connection on Q is an assignment to each pair of vector elds X and Y on Q of a vector eld X Y , where the assignment satises: (i) (X, Y ) (ii) (iii) i, j, k {1, . . . , n}.
fX Y XY XY
d L L i i dt v q
= Gij q + G
G
jk
= Gij q j + j q l q m , lm where i jk
G
is R-bilinear; for all vector elds X and Y , and all functions f ; + (L X f )Y for all vector elds X and Y , and all functions
=f )=f
Glj 1 Glk Gjk , = Gil + k j 2 q q q l

G
X (f Y
XY
f. The vector eld

XY
Question: What are these functions i ? jk
is the covariant derivative of Y with respect to X.
Ane connections (contd) Ane connections (contd) Question: What really characterizes ? A connection is completely determined by its Christoel symbols:
XY
Coordinate answer: Let (q 1 , . . . , q n ) be coordinates. Dene n3 functions i , jk i, j, k {1, . . . , n}, on the chart domain by
q j
Y i j X + i X j Y k . jk j q q i
Theorem 12. Let G be a Riemannian metric on a manifold Q. Then there exists

G
= i , jk k q q i
a unique ane connection

G
, called the Levi-Civita connection, such that

G
j, k {1, . . . , n}. in the given
(i) L X (G(Y, Z)) = G(

G G
X Y, Z)
+ G(Y,
G
X Z)
and
i , i, j, k {1, . . . , n}, are the Christoel symbols for jk coordinates.
(ii)
XY
Furthermore, the Christoel symbols of
X = [X, Y ].
G
are i , i, j, k {1, . . . , n}. jk
Return to EulerLagrange equations Had shown that
Forces q +
i
d L L i =0 i dt v q
i q j q k jk
= 0.
Some linear algebra: If V is a R-vector space, V is the set of linear maps from V to R. This is the dual space of V. Denote (v) = ; v for V and v V. If {e1 , . . . , en } is a basis for V, the dual basis for V is denoted by {e1 , . . . , en } i and dened by ei (ej ) = j . The dual space of Tq Q is denoted by T Q, and called the cotangent space. q
The dual basis to { q1 , . . . , qn } is denoted by {dq 1 , . . . , dq n }.
Interpretation of q i + i q j q k . jk
(i) Covariant derivative of with respect to itself: i j k qi (t) (t) = ( + jk q q ) q i .
(ii) Curves t (t) satisfying (t) (t) = 0 are geodesics and can be thought of as being acceleration free.
G
(iii) Mechanically,
G
(t) (t) = 0 . accn

force mass
A covector eld assigns to each point q Q an element of T Q. q
Bottom-line: (t) (t) can be computed, and gives access to signicant mathematical tools.
Example 13. The dierential of a function is df (q) T Q dened by q f df (q); X(q) = L X f (q). In coordinates, df = qi dq i .
Forces (contd) Newtonian forces on a rigid body: force f applied to the center of mass and a pure torque . Need to add these to the EulerLagrange equations in the right way. Use the idea of innitesimal work done by a (say) force f in the direction w: f , w R3 . For torques, the analogue is , the angular velocity.
R3
Forces (contd) Fix body j with Newtonian force f j and torque j . Let t (t) satisfy (0) = wq , and let t (Rj (t), r j (t)) = j (t). Let j (t) = Rj (t)R1 (t) be the spatial angular velocity. j Dene Ff j , j T Q by q Ff j , j ; wq = f j , r j (0)
R3
where is the spatial representation of
+ j , j (0)
R3
.
k j=1
Interconnected mechanical system with conguration manifold Q, q Q, wq Tq Q. Determine force as element of T Q by its action on wq . q
Sum over all bodies to get total external force F T Q: F = q
Ff j , j .
Forces (contd) Note that the forces may depend on time (e.g., control forces) and velocity (e.g., dissipative forces). A force is a map F : R TQ T Q satisfying F (t, vq ) T Q. q Thus can write F = Fi (t, q, v)dq . Question: How do forces appear in the EulerLagrange equations? Answer: Like this: d L L i = Fi . i dt v q Why? Because this agrees with Newton.
i
Forces (contd) Given a force F : R TQ T Q, dene a vector force G (F ) : R TQ TQ by G(G (F )(t, vq ), wq ) = F (t, vq ); wq .
In coordinates, G (F ) = Gij Fj qi .
The EulerLagrange equations subject to force F are then equivalent to

G
(t) accn
(t) = G (F )(t, (t))

force mass
F1
b2
b2,2
Forces (contd)
F2
b1
Forces (contd)
b2,1 2 b1,2 b1,1 s2 1 s1 ag F2
Example 14. Planar rigid body: Example 15. Two-link manipulator: f 1,1 = F (cos( + ), sin( + ), 0), 1,1 = F (0, 0, h sin ), f 2,1 = (0, 0, 0), F 2 = d. 2,1 = (0, 0, 1), F 1 = F cos( + )dx + sin( + )dy h sin d , Equations of motion easily computed.
F1
1,1 = 1 (0, 0, 1), 1,2 = (0, 0, 0), 2,1 = 2 (0, 0, 1), 2,2 = 2 (0, 0, 1), F 1 = 1 d1 , F 2 = 2 (d2 d1 ). Gravitational force and equations of motion easily computed.
F1
Forces (contd)
b2 s3 b3 s2 s1 F2 b1
Distributions and codistributions A distribution (smoothly) assigns to each point q Q a subspace Dq of Tq Q. A codistribution (smoothly) assigns to each point q Q a subspace q of T Q. q We shall always consider the case where the function q dim(Dq ) (resp. q dim(q )) is constant, although there are important cases where this does not hold. Given a distribution D, dene a codistribution ann(D) by ann(D)q = { q | q (vq ) = 0 for all vq Dq }. Given a codistribution , dene a distribution coann() by coann()q = { vq | q (vq ) = 0 for all q q }.
Example 16. Rolling disk: 1,1 = 1 (0, 0, 1), 2,1 = 2 ( sin , cos , 0), F 1 = 1 d, F 2 = 2 d.
Equations of motion cannot be computed yet, because we have not dealt with. . . nonholonomic constraints.
Nonholonomic constraints An interconnected mechanical system with conguration manifold Q, kinetic energy metric G and external force F . A nonholonomic constraint restricts the set of admissible velocities at each point q to lie in a subspace Dq , i.e., it is dened by a distribution D. Example 17. At a conguration q with coordinates (x, y, , ), the admissible velocities satisfy
s3
Nonholonomic constraints (contd) Question: What are the equations of motion for a system with nonholonomic constraints?
b2 b3
b1
Answer: Determined by the LagrangedAlembert Principle. We will skip a lot of physics and metaphysics, and go right to the ane connection formulation, originally due to Synge [1928].
x = cos
s2
y = sin . Thus Dq has {X1 (q), X2 (q)} as basis, where X1 = cos
s1 (x, y)
+ sin + , x y
X2 =
Ane connection control systems Control force assumption: Directions in which control forces are applied depend only on position, and not on time or velocity. There exists covector elds F 1 , . . . , F m such that the control force takes m the form Fcon = a=1 ua F a . Control forces appear in equations of motion after application of G and (possibly) projection by PD . Model eects of input forces by vector elds Y1 , . . . , Ym . Model uncontrolled external forces by vector force Y . Nothing to be gained by assuming that ane connection comes from physics. Use arbitrary ane connection . Control equations:
m (t)
Nonholonomic constraints (contd) Let D be the G-orthogonal complement to D, let PD be the G-orthogonal projection onto D, and let PD be the G-orthogonal projection onto D .
D
Dene an ane connection

D
by
G G
XY
XY
+(
X PD )(Y
).
Theorem 18. The following are equivalent: (i) t (t) is a trajectory for the system subject to the external force F ;
D
(ii)
(t)
(t) = PD (G (F )(t, (t))).
(t) =
a=1
ua (t)Ya ((t)) + Y (t, (t)),
Ane connection control systems (contd) Denition 19. A forced ane connection control system is a 6-tuple = (Q, , D, Y, Y = {Y1 , . . . , Ym }, U ) where (i) Q is a manifold, (ii) is an ane connection such that in D,
XY
Ane connection control systems (contd) Denition 20. A control-ane system is a triple = (M, C = {f0 , f1 , . . . , fm }, U ) where (i) M is a manifold, takes values in D if Y takes values (ii) f0 , f1 , . . . , fm are vector elds on M, and (iii) U Rm . Control equations:
m
(iii) D is a distribution, (iv) Y is a vector force taking values in D, (v) Y1 , . . . , Ym are D-valued vector elds, and (vi) and U Rm . Take away forced if Y = 0.
(t) = f0 ((t)) +
drift vector eld a=1
ua (t) fa ((t)) .
control vector eld
Ane connection control systems (contd) Ane connection control systems are control-ane systems. (i) The state manifold is M = TQ. Representations of control equations Global representation:
m (t) (t) = a=1
(ii) The drift vector eld is denoted by S and called the geodesic spray. Coordinate expression: i v j v k i f0 = S = v jk q i v
i
cf. q +
i q j q k jk
=0 .
ua (t)Ya ((t)) + Y (t, (t)).
(iii) The control vector elds are the vertical lifts vlft(Ya ) of the vector elds Ya , a {1, . . . , m}. Coordinate expression:
i fa = vlft(Ya ) = Ya i . v
Natural local representation:

m
q i + i q j q k = jk
a=1
i ua Ya + Y i ,
i {1, . . . , m}.
Can add external force to drift to accommodate forced ane connection control systems.
Representations of control equations (contd) Let X = {X1 , . . . , Xn } be vector elds dened on a chart domain U with the property that, for each q U, {X1 (q), . . . , Xn (q)} is a basis for Tq Q. For q U and wq Tq Q, write wq = v i Xi (q); {v 1 , . . . , v n } are pseudo-velocities. ua (t)vlft(Ya )((t)).
a=1
Representations of control equations (contd) Global rst-order representation:

m
(t) = S((t)) + vlft(Y )((t)) + Natural rst-order local representation: qi = vi ,

m
The generalized Christoel symbols are

Xj X k
= i Xi , jk
j, k {1, . . . , n}.
i {1, . . . , n},
i ua Ya ,
Poincar local representation: e

i q i = Xj v j ,
X
v i = i v j v k Y i + jk
a=1
i {1, . . . , n}.
i {1, . . . , n}, i ua Ya , i {1, . . . , n},
vi = i vj vk Y i + jk
a=1
where means components with respect to the basis X .
Representations of control equations (contd)

D
In the case when = , this simplies when we choose {X1 , . . . , Xn } such that {X1 (q), . . . , Xk (q)} forms a G-orthogonal basis for Dq .
X
(q) =
1 X (q)
2 G( G
X X (q), X (q)),
, , {1, . . . , k}.
Representations of control equations (contd) Seems unspeakably ugly, but is easily automated in symbolic manipulation language. Snakeboard example.
Signicant advantages in symbolic computation. orthogonal Poincar representation: e

i q i = X v ,
X
v = v v +
1 X
2 G
i {1, . . . , n}, ua F a ; X , {1, . . . , k}.
F ; X +
a=1
Reachable sets for control-ane systems
Controllability theory
(i) Denitions of controllability and background for control-ane systems (ii) Accessibility theorem (iii) Controllability denitions and theorems for ACCS (iv) Good/bad conditions (v) Examples (vi) Snakeboard using Mma (vii) Series expansions
A control-ane system = (M, C = {f0 , f1 , . . . , fm }, U ) A controlled trajectory of is a pair (, u), where u : I U is locally integrable, and : I M is the locally absolutely continuous
m
(t) = f0 ((t)) +
a=1
ua (t)fa ((t))
Ctraj(, T ) is set of controlled trajectories (, u) for dened on [0, T ] Dene the various sets of points that can be reached by trajectories of a control-ane system. For x0 M, the reachable set fof from x0 is R (x0 , T ) = { (T ) | (, u) Ctraj(, T ), (0) = x0 }
Controllability notions for control-ane systems = (M, C = {f0 , f1 , . . . , fm }, U ) is C -control-ane system, x0 M Involutive closure D is a smooth distribution if it has smooth generators a distribution is involutive if it is closed under the operation of Lie bracket inductively dene distributions Lie(l) (D), l {0, 1, 2, . . . } by Lie(0) (D)x = Dx Lie(l) (D)x = Lie(l1) (D)x + span [X, Y ](x) X takes values in Lie(l1 ) (D) Y takes values in Lie(l2 ) (D),
x0 R (x0 , T ) R (x0 , T ) R (x0 , T ) x0 x0
is accessible from x0 if there exists T > 0 such that int(R (x0 , t)) = for t ]0, T ] is controllable from x0 if, for each x M, there exists a T > 0 and (, u) Ctraj(, T ) such that (0) = x0 and (T ) = x is small-time locally controllable (STLC) from x0 if there exists T > 0 such that x0 int(R (x0 , t)) for each t ]0, T ]
l1 + l 2 = l 1
the involutive closure Lie() (D) is the pointwise limit
not accessible
accessible
controllable (STLC)
Theorem 21. (Under smoothness and regularity assumptions) Lie() (D) contains D and is contained in every involutive distribution containing D
Examples of accessible control-ane systems
Accessibility results for control-ane systems = (M, C , U ) is an analytic control-ane system we say satises the Lie algebra rank condition (LARC) at x0 if Lie
()
s3 s2 s1
b2 b3
b1
(x, y)
(C )x0 = Tx0 M
rank Lie
()
(C )x0 = n x cos 0 y sin 0 = u1 + u2 0 1 1 0
a control set U is proper if 0 int(conv(U ))
Theorem 22. If U is proper, then is accessible from x0 if and only if satises LARC at x0 It is not known if there are useful necessary and sucient conditions for STLC. Available results include a sucient condition given as the neutralization of bad bracket by good brackets of lower order
(unicycle dynamics, simplest wheeled robot dynamics)
xr 0 cos y sin 0 r = 1 u1 + u2 tan 0 1 0
Trajectories and reachable sets of mechanical systems (time-independent) general simple mechanical control system = (Q, G, V, F, D, F = {F 1 , . . . , F m }, U ) a controlled trajectory for is pair (, u), with u : I U and : I Q, satisfying (t0 ) D(0t ) for some t0 I and Summary notions of accessibility and STLC tool: Lie bracket and involutive closure necessary and sucient conditions for conguration accessibility reachable sets from states with zero velocity: R,TQ (q0 , T ) = { (T ) | (, u) Ctraj(, T ), (0) = 0q0 } , R,Q (q0 , T ) = { (T ) | (, u) Ctraj(, T ), (0) = 0q0 } ,
t[0,T ]
D
(t)
(t) = PD (gradV ((t))) + PD (G (F ( (t))))

m
+
a=1
ua (t)PD (G (F a ((t)))).
Ctraj(, T ) is set of [0, T ]-controlled trajectories for on Q
R,TQ (q0 , T ) =
R,TQ (q0 , t),
R,Q (q0 , T ) =
t[0,T ]
R,Q (q0 , t).
Controllability notions for mechanical systems
Controllability for mechanical systems: linearization results Let = (Rn , M , K, F ) be a linear mechanical control system, i.e., M and K are square n n matrices and F is n m,
= (Q, G, V, F, D, F , U ) is general simple mechanical control system with F time-independent, U proper, and q0 Q is accessible from q0 if there exists T > 0 such that intD (R,TQ (q0 , t)) = for t ]0, T ] is conguration accessible from q0 if there exists T > 0 such that int(R,Q (q0 , t)) = for t ]0, T ] is small-time locally controllable (STLC) from q0 if there exists T > 0 such that 0q0 intD (R,TQ (q0 , t)) for t ]0, T ]. is small-time locally conguration controllable (STLCC) from q0 if there exists T > 0 such that q0 int(R,Q (q0 , t)) for t ]0, T ].
M x(t) + Kx(t) = F u(t) Theorem 23. The following two statements are equivalent: (i) is STLC from 0 0 (ii) the following matrix has maximal rank M 1 F M 1 K (M 1 F ) (M 1 K)n1 (M 1 F )
Corresponding linearization result where, in coordinates, M = G(q0 ), K = Hess V (q0 ), and no dissipation Corollary 24. If = (Q, G, V = 0, F , U ) is underactuated at q0 , then its linearization about 0q0 is not accessible from the origin.
Symmetric product as a Lie bracket Given vector eld Y on Q, its vertical lift vlft(Y ) is vector eld on TQ Y1 0 . vlft(Y ) = Y i i = 0 Y Y = Y i i . , q v . Y Yn Recall: The drift vector eld S and called the geodesic spray: S = vi In coordinates X:Y
k
The symmetric product given manifold Q with ane connection the symmetric product corresponding to is the operation that assigns to vector elds X and Y on Q the vector eld X:Y =
XY
i v j v k i jk i q v
Y k i X k i X + Y + k X i Y j + X j Y i ij q i q i
remarkable Lie bracket identities: [S, vlft(Y )](0q ) = Y (q) 0q [vlft(Ya ), [S, vlft(Yb )]](vq ) = vlft( Ya : Yb )(vq )
Symmetric closure take smooth input distribution Y a distribution is geodesically invariant if it is closed under the operation of symmetric product inductively dene distributions Sym (Y), l {0, 1, 2, . . . } by Sym
(0) (l)
Accessibility results for mechanical systems = (Q, , D, Y = {Y1 , . . . , Ym }, U ) is an analytic ACCS U proper q0 point in Q Theorem 26. (i) is accessible from q0 if and only if Sym() (Y)q0 = Dq0 and Lie() (D)q0 = Tq0 Q (ii) is conguration accessible from q0 if and only if Lie() (Sym() (Y))q0 = Tq0 Q Key result in proof: If C = {S, vlft(Y1 ), . . . , vlft(Ym )}, then, for q0 Q, Lie() (C )0q0 Lie() (Sym() (Y))q0 Sym() (Y)q0
(Y)q = Yq X takes values in Sym(l1 ) (D) Y takes values in Sym(l2 ) (D), l1 + l 2 = l 1

()
Sym(l) (Y)q = Sym(l1) (Y)q + span X : Y (q)
the symmetric closure Sym
()
(Y) is the pointwise limit
Theorem 25. (Under smoothness and regularity assumptions) Sym (Y) contains Y and is contained in every geodesically invariant distribution containing Y
Controllability mechanisms Notions for sucient test
given control forces {F 1 , . . . , F m }
Consider iterated symmetric products in the vector elds {Y1 , . . . , Ym }: (i) A symmetric product is bad if it contains an even number of each of the vector elds Y1 , . . . , Ym , and otherwise is good. E.g., Ya : Yb : Ya : Yb is bad, Ya : Yb : Yc is good
accessible accelerations {Y1 , . . . , Ym } Ya = PD (G1 F a )
(ii) The degree of a symmetric product is the total number of input vector elds comprising the symmetric product. E.g., Ya : Yb : Ya : Yb has degree 4
access. velocities Sym() (Y1 , . . . , Ym ) {Yi , Yj : Yk , Yj : Yk : Yh , . . . }
decoupling v.f.s {V1 , . . . , V } Vi , Vi : Vi {Y1 , . . . , Ym }
(iii) If P is a symmetric product and if is a permutation on {1, . . . , m}, dene (P ) as symmetric product where each Ya is replaced with Y(a)
access. confs Lie() (Sym() (Y1 , . . . , Ym )) {Yi , Yj : Yk , [Yj , Yk ], [ Yj : Yk , Yh ], . . . }
Lie() (V1 , . . . , V ): congurations accessible via decoupling v.f.s
Controllability for ACCS ACCS = (Q, , D, Y, U ), q0 Q, U proper satises bad vs good condition if for every bad symmetric product P
Sm
Summary for control-ane systems notions of accessibility and STLC tool: Lie bracket and involutive closure necessary and sucient conditions for accessibility Summary for ACCS notions of conguration accessibility and STLCC tool: symmetric product and symmetric closure necessary and sucient conditions for accessibility
(P )(q0 ) spanR {P1 (q0 ), . . . , Pk (q0 )}
where P1 , . . . , Pk are good symmetric products of degree less than P Theorem 27. rank Sym() (Y)q0 is maximal bad vs good STLC= small-time locally controllable (q0 , 0) (qf , vf ) can reach open set of congurations and velocities STLCC= small-time locally conguration controllable (q0 , 0) (qf , vf ) of congurations
u u
rank Lie() (Sym() (Y))q0 = n bad vs good
can reach open set
Controllability examples

F1
Y1 is internal torque and Y2 is extension force. Both inputs: not accessible, conguration accessible, and STLCC (satises sufcient condition). Y1 only: conguration accessible but not STLCC. Y2 only: not conguration accessible.
Y1 is component of force along center axis, and Y2 is component of force perpendicular to center axis. Y1 and Y2 : accessible and STLCC (satises sucient condition).
F2
Y1 and Y3 : accessible and STLCC (satises sucient condition). Y1 only or Y3 only: not conguration accessible. Y2 only: accessible but not STLCC. Y2 and Y3 : conguration accessible and STLCC (but fails sucient condition).
Y1 is rolling input and Y2 is spinning input. Y1 and Y2 : conguration accessible and STLCC (satises sucient condition). Y1 only: not conguration accessible. Y2 only: not conguration accessible.
s3 s2 s1 (x, y)
Y1 rotates wheels and Y2 rotates rotor.
Y1 and Y2 : conguration accessible and STLCC (satises sucient condition). Y1 only: not conguration accessible. Y2 only: not conguration accessible.
Series expansion for ane connection control systems = (Q, , D, Y = {Y1 , . . . , Ym }, U ) is an analytic ACCS
(t)
(t) = Y (t, (t)) (0) = 0
(t) =
k=1
Vk (t, (t))
absolute, uniform convergence
Single input at joint. Conguration accessible, but not STLCC.
V1 (t, q) =
0
Y (s, q)ds 1 2
k1 0 t
Vk (t, q) =
Vj (s, q) : Vkj (s, q) ds

j=1
Series: comments
+
(t) =
k=1
Vk (t, (t))
V1 (t, q) =
Error bounds:
Vk+1 (t, q) = Vk = O( Y
t Y 0 1 2
(s, q)ds
t 0
Va (s, q) : Vka (s, q) ds
Kinematic reductions and motion planning

(i) Motion planning problems for driftless systems and ACCS (ii) How to reduce the MPP for ACCS to the MPP for a driftless system (iii) Kinematic reductions: notion, theorems and examples
k 2k1
In abbreviated notation V1 = Y , so that (t) = Y (t, (t)) 1 1 Y : Y (t, (t)) + 2 2 Y : Y : Y (t, (t)) + O( Y
4 7
(iv) Kinematic controllability V2 = 1 Y :Y , 2 V3 = 1 2 Y :Y :Y (v) Inverse kinematics and example solutions (vi) Motion planning problems with animations
t )
Motion planning for driftless systems: contd Examples of U -motion planning problem Motion planning for driftless systems (M, {X1 , . . . , Xm }, U ) is driftless system:
m
(i) motion planning problem with continuous inputs U = {e1 , . . . , em , e1 , . . . , em } U is collection of piecewise constant U -valued functions Then, is concatenation of integral curves, possibly running backwards in time, of the vector elds X1 , . . . , Xm . Each curves is a primitive
(ii) motion planning problem using primitives:
(t) =
a=1
Xa ((t))ua (t)
where u are U -valued integrable inputs let U be a set of inputs U -motion planning problem is:
Given x0 , x1 M, nd u U , dened on some interval [0, T ], so that the controlled trajectory (, u) with (0) = x0 satises (T ) = x1
Motion planning using primitives Consider (M, {X1 , . . . , Xm }, Rm ). If Lie() (X ) = TM, then, for each x0 , x1 M, there exist k N, t1 , . . . , tk R, and a1 , . . . , ak {1, . . . , m} such that x1 = tk
Xa k
t1a1 (x0 )
Technical conditions: smoothness, complete vector elds, M connected
How to reduce the MPP for ACCS to the MPP for a driftless system Motion planning for ACCS (Q, , D, {Y1 , . . . , Ym }, U ) is ane connection control system (ACCS)
m (t)
Key idea: Kinematic Reductions Goal: (low-complexity) kinematic representations for mechanical control systems Consider an ACCS, i.e., systems with no potential energy, no dissipation (i) ACCS model with accelerations as control inputs mechanical systems:
m (t)
(t) =
a=1
u (t)Ya ((t))
U is set of U -valued integrable inputs U -motion planning problem is: Given q0 , q1 Q, nd u U , dened on some interval [0, T ], so that the controlled trajectory (, u) with (0) = 0q0 has the property that (T ) = 0q1
(t) =
a=1
Ya ((t))ua (t)
Y = span {Y1 , . . . , Ym }
(ii) driftless = kinematic model with velocities as control inputs (t) =

b=1
Vb ((t))wb (t)
V = span {V1 , . . . , V }
is the rank of the reduction
When can a second order system follow the solution of a rst order? Kinematic reductions ex: Can follow any straight line and can turn 2 preferred velocity elds (plus, conguration controllability) V = span {V1 , . . . , V } is a kinematic reduction if any curve q : I Q solving the (controlled) kinematic model can be lifted to a solution of the (controlled) dynamic model. rank 1 reductions are called decoupling vector elds The kinematic model induced by {V1 , . . . , V } is a kinematic reduction of (Q, , D, {Y1 , . . . , Ym }, U ) if and only if
" ! " ! #
Ok
(i) V Y (ii) V : V Y

" !
"

% $
&
Examples of kinematic reductions

Three link planar manipulator with passive link
Two rank 1 kinematic reductions (decoupling vector elds) no rank 2 kinematic reductions
Actuator conguration (0,1,1) (1,0,1) (1,1,0)
Decoupling vector elds 2 2 2
Kinematically controllable yes yes yes
Examples of maximally reducible systems
When is a mechanical system kinematic? When are all dynamic trajectories executable by a single kinematic model? A dynamic model is maximally reducible (MR) if all its controlled trajectory (starting from rest) are controlled trajectory of a single kinematic reduction. (Q, , D, {Y1 , . . . , Ym }, U ) is maximally reducible if and only if (i) the kinematic reduction is the input distribution Y (ii) Y : Y Y
b2 b3
b1
s3 s2 s1 (x, y)

(unicycle dynamics, simplest wheeled robot dynamics)
x cos 0 y sin 0 = v + 0 1 1 0
xr 0 cos y sin 0 r = 1 v + tan 0 0 1
Kinematic controllability Objective: controllability notions and tests for mechanical systems and reductions Consider: (Q, , D, {Y1 , . . . , Ym }, U ) KC= locally kinematically controllable V1 , . . . , V decoupling v.f.s rank Lie() (V1 , . . . , V ) = n (q0 , 0) (qf , 0) can reach open set of congurations by concatenating motions along kinematic reductions STLC= small-time locally controllable (q0 , 0) (qf , vf ) can reach open set of congurations and velocities STLCC= small-time locally conguration controllable (q0 , 0) (qf , vf ) of congurations
u u u
Controllability mechanisms
given control forces {F 1 , . . . , F m }
accessible accelerations {Y1 , . . . , Ym } Ya = PD (G1 F a )
rank Sym() (Y) = n, bad vs good
access. velocities Sym() (Y1 , . . . , Ym ) {Yi , Yj : Yk , Yj : Yk : Yh , . . . }
decoupling v.f.s {V1 , . . . , V } Vi , Vi : Vi {Y1 , . . . , Ym }
rank Lie (Sym bad vs good
()
()
(Y)) = n,
can reach open set
access. confs Lie() (Sym() (Y1 , . . . , Ym )) {Yi , Yj : Yk , [Yj , Yk ], [ Yj : Yk , Yh ], . . . }
Lie() (V1 , . . . , V ): congurations accessible via decoupling v.f.s
Controllability inferences STLC STLCC KC MR-KC = = = = small-time locally controllable small-time locally conguration controllable locally kinematically controllable maximally reducible, locally kinematically controllable
Cataloging kinematic reductions and controllability of example systems

System planar 2R robot single torque at either joint: (1, 0), (0, 1) n = 2, m = 1 roller racer single torque at joint n = 4, m = 1 Picture Reducibility (1, 0): no reductions (0, 1): maximally reducible Controllability accessible not accessible or STLCC
no kinematic reductions
accessible, not STLCC
STLC
KC
MR-KC
planar body with single force or torque n = 3, m = 1 planar body with single generalized force n = 3, m = 1 planar body with two forces n = 3, m = 2
decoupling v.f.
reducible, not accessible
STLCC
no kinematic reductions
accessible, not STLCC
two decoupling v.f.
KC, STLC
There exist counter-examples for each missing implication sign.
robotic leg n = 3, m = 2
two decoupling v.f., maximally reducible (1, 0, 1) and (1, 1, 0): two decoupling v.f. (0, 1, 1): two decoupling v.f. and maximally reducible fully reducible
KC
planar 3R robot, two torques: (0, 1, 1), (1, 0, 1), (1, 1, 0) n = 3, m = 2 rolling penny n = 4, m = 2
(1, 0, 1) and (1, 1, 0): KC and STLC (0, 1, 1): KC KC
Summary relationship between trajectories of dynamic and of kinematic models of mechanical systems kinematic reductions (multiple, low rank), and maximally reducible systems
snakeboard n = 5, m = 2 3D vehicle with 3 generalized forces n = 6, m = 3
two decoupling v.f.
KC, STLCC
controllability mechanisms, e.g., STLC vs kinematic controllability
three decoupling v.f.
KC, STLC
Left invariant vector elds on matrix Lie groups Matrix Lie groups are manifolds of matrices closed under the operations of matrix multiplication and inversion Example: SO(3) = R R33 RRT = I3 , det(R) = +1
Trajectory design via inverse kinematics Objective: nd u such that (qinitial , 0) (qtarget , 0) Assume: (i) (Q, , D, {Y1 , . . . , Ym }, U ) is kinematically controllable (ii) Q = G and decoupling v.f.s {V1 , . . . , V } are left-invariant
u
left invariant vector elds have the following properties: (i) R(t) = XV (R(t)) = R(t) V for some matrix V XV (R0 ) = R0 exp(tV ) t
(linear dependence)
(ii) ow of left invariant vector eld is equal to left multiplication
(iii) exp(tV ) SO(3), that is, V so(3) set of skew symmetric matrices (iv) For e1 , e2 , e3 the standard basis of R3 , 0 0 0 0 0 0 0 1 , e2 = 0 0 e1 = 0 1 0 1 0 1 0 1 1 0 e3 = 0 0 0
0 , 0
0 0
Trajectory design via inverse kinematics Objective: nd u such that (qinitial , 0) (qtarget , 0) Assume: (i) (Q, , D, {Y1 , . . . , Ym }, U ) is kinematically controllable (ii) Q = G and decoupling v.f.s {V1 , . . . , V } are left-invariant = matrix exponential exp : g G gives closed-form ow = composition of ows is matrix product Objective: select a nite-length combination of k ows along {V1 , . . . , V } and coasting times {t1 , . . . , tk } such that
1 qinitial qtarget u
Inverse-kinematic planner on SO(3) Any underactuated controllable system on SO(3) is equivalent to V1 = ez = (0, 0, 1) V2 = (a, b, c) with a2 + b2 = 0
Motion Algorithm: given R SO(3), ow along (ez , V2 , ez ) for coasting times t1 = atan2 (w1 R13 + w2 R23 , w2 R13 + w1 R23 ) t2 = acos R33 c2 1 c2
t3 = atan2 (v1 R31 + v2 R32 , v2 R31 v1 R32 )

where z = 4 2 1 cos t2 sin t2 3 5, 2 4 w1 w2 3 5=4 2 ac cb b a 3 5 z,
FK
= gdesired = exp(t1 Va1 ) exp(tk Vak ).
2 3 2 ac v 4 15 = 4 v2 cb
b a
5z
No general methodology is available = catalog for relevant example systems SO(3), SE(2), SE(3), etc
Local Identity Map =
R (t1 , t2 , t3 ) exp(t1 ez ) exp(t2 V2 ) exp(t3 ez )
IK
Inverse-kinematic planner on SO(3): simulation The system can rotate about (0, 0, 1) and (a, b, c) = (0, 1, 1) Rotation from I3 onto target rotation exp(/3, /3, 0) As time progresses, the body is translated along the inertial x-axis First class of underactuated controllable system on SE(2) is 1 = {(V1 , V2 )| V1 = (1, b1 , c1 ), V2 = (0, b2 , c2 ), b2 + c2 = 1} 2 2 Motion Algorithm: given (, x, y), ow along (V1 , V2 , V1 ) for coasting times (t1 , t2 , t3 ) = (atan2 (, ) , , atan2 (, ))
2 3 2 p b2 where = 2 + 2 and 4 5 = 4 c2 c2 3 02 3 2 x c 5 @4 5 4 1 b2 y b1
FK
Inverse-kinematic planner for 1 -systems SE(2)
32 31 1 cos 54 5A c1 sin b1
Identity Map =
(, x, y) (t1 , t2 , t3 ) exp(t1 V1 ) exp(t2 V2 ) exp(t3 V1 )
IK
Inverse-kinematic planner for 2 -systems SE(2) Second and last class of underactuated controllable system on SE(2): 2 = {(V1 , V2 )| V1 = (1, b1 , c1 ), V2 = (1, b2 , c2 ), b1 = b2 or c1 = c2 } Motion Algorithm: given (, x, y), ow along (V1 , V2 , V1 ) for coasting times t1 = atan2 , 4 2 + atan2 (, ) t3 = t1 t2
2 3 2 p c1 c2 where = 2 + 2 , 4 5 = 4 b1 b2 3 02 3 2 x c 5 @4 5 4 1 c1 c2 y b1 b2 b1
FK
Inverse-kinematic planners on SE(2): simulation
t2 = atan2 2 2 , 4 2
32 54 31 5A
b1 c1
1 cos sin
Local Identity Map =
(, x, y) (t1 , t2 , t3 ) exp(t1 V1 ) exp(t2 V2 ) exp(t3 V1 )
IK
Inverse-kinematics planners for sample systems in 1 and 2 . The systems parameters are (b1 , c1 ) = (0, .5), (b2 , c2 ) = (1, 0). The target location is (/6, 1, 1).
Inverse-kinematic planners on SE(2): snakeboard simulation
Inverse-kinematic planners on SE(2) R: simulation 4 dof system in R3 , no pitch no roll kinematically controllable via body-xed constant velocity elds: V1 = rise and rotate about inertial point; V2 = translate forward and dive
snakeboard as 2 -system
The target location is (/6, 10, 0, 1)
Inverse-kinematic planners on SE(3): simulation kinematically controllable via body-xed constant velocity elds: V1 = translation along 1st axis V2 = rotation about 2nd axis V3 = rotation about 3rd axis
5 yg 4 xg zg
Summary relationship between trajectories of dynamic and of kinematic models of mechanical systems kinematic reductions (multiple, low rank), and maximally reducible systems controllability mechanisms, e.g., STLC vs kinematic controllability systems on matrix Lie groups inverse-kinematics planners
V3 V2 V1 V3 V2 V3
: 0 1: : 1 2: : 2 3: : 3 4: : 4 5: : 5 6:
rotation about 3rd axis rotation about 2nd axis translation along 1st axis rotation about 3rd axis rotation about 2nd axis rotation about 3rd axis
6 z0 y0 3
x0
Analysis and design of oscillatory controls for ACCS

(i) Introduction to Averaging (ii) Survey of averaging results (iii) Two-time scale averaging analysis for mechanical systems (iv) Analysis via the Averaged Potential (v) Control design via Inversion Lemma (vi) Tracking results and examples
Introduction to averaging Oscillations play key role in animal and robotic locomotion oscillations generate motion in Lie bracket directions useful for trajectory design objective is to study oscillatory controls in mechanical systems:
T (t)
(t) = Y (t, (t)),

0
Y (t, (t))dt = 0
oscillatory signals: periodic large-amplitude, high-frequency
Averaging for systems in standard form Survey of results on averaging Early developments: Lagrange, Jacobi, Poincar e Oscillatory Theory: Dynamical Systems: Bogoliubov Mitropolsky, Guckenheimer Holmes, Sanders Verhulst, . . . for > 0, system in standard form (t) = X(t, (t)), (0) = x0
assume X is T -periodic, dene the averaged vector eld X(x) = 1 T

T
Control Systems: Bloch, Khalil . . . Related Work:
X(, x)d.
0
General ODEs: Kurzweil-Jarnik, Sussmann-Liu,
dene the averaged trajectory t (t) M by (t) = X((t)), Theorem 28 (First-order Averaging Theorem). (t) (t) = O() for all t [0, t0 ] (0) = x0
(Electro)Mechanical Systems: Hill, Mathieu, Bailleiul, Kapitsa, Levi . . . Series Expansions: Magnus, Chen, Brockett, Gilbert, Sussmann, Kawski . . . Time-dependent vector elds: Agrachev, Gramkrelidze, . . . Small-amplitude averaging and high-order averaging: Sarychev, Vela, . . .
If X has linearly asymptotically stable point, then estimate holds for all time
Averaging for systems in standard oscillatory form for > 0, system in standard oscillatory form 1 t , t, (t) , (t) = X(t, (t)) + Y Assumptions: (0) = x0
Averaging for systems in standard oscillatory form: contd dene F as average with respect to for xed 0 , compute the trajectories (t) = F (t, (t)) (t, 0 ) = Y (t, 0 , (t)) with initial conditions: (0) = x0 and (0) = (t) (note (, t) equals (t) plus zero-mean oscillation) (t) (t/, t) = O() for all t [0, t0 ]
(i) Y is T -periodic and zero-mean in rst argument
(ii) the vector elds x Y (, t, x), at xed (, t), are commutative Useful constructions: (i) given dieomorphism and vector eld X, the pull-back vector eld X = T 1 X
(ii) given extended state xe = (t, x), dene Xe (xe ) = (1, X(xe )), and Ye (, xe ) = (0, Y (, xe )) (iii) dene F as two-time scale vector eld by (1, F (, xe )) = (Ye ) Xe (xe ) 0,
Theorem 29 (Oscillatory Averaging Theorem).
Two-time scale averaging for mechanical systems for R+ , consider the forced ACCS (Q, , Y, D, Y = {Y1 , . . . , Ym }, Rm ):
(t)
dene the averaged ACCS

m (t) (t) = Y (t, (t))
ab (t) Ya : Yb ((t))
a,b=1
(t) = Y (t, (t)) +
1 a t u , t Ya ((t)) a=1
with initial condition

m
where Y is an ane map of the velocities assume the two-time scale inputs u = (u , . . . , u ) : R+ R+ Rm are T -periodic and zero-mean in their rst argument
1 m
(0) = (0) +
a=1
U (a) (0)Ya ((0))
dene the symmetric positive-denite curve : R+ Rmm by ab (t) = where

1 2
Theorem 30 (Oscillatory Averaging Theorem for ACCS). there exists 0 , t0 R+ such that, for all t [0, t0 ] and for all (0, 0 ), (t) = (t) + O(),
m
U(a) U(b) (t) U (a) (t)U (b) (t) , 1 T
a, b {1, . . . , m}
T
(t) = (t) +
a=1
t U(a) ( , t) U (a) (t) Ya ((t)) + O().
U(a) (, t) =
0
ua (s, t)ds,
U (a) (t) =
U(a) (, t)d
0
If oscillatory inputs depend only on fast time, and if the averaged ACCS has linearly asymptotically stable equilibrium conguration, then estimate holds for all time
Averaging analysis with potential control forces when is the averaged system again a simple mechanical system? consider simple mechanical control system (Q, G, V, Fdiss , F , R ) (i) no constraints
G
Averaging via the averaged potential
(iii) Fdiss is linear in velocity dene input vector elds
(ii) F = {d1 , . . . , dm }, where a : Q R for a {1, . . . , m}
(t)
(t) = gradV ((t)) + G (Fdiss ( (t))) + 1 a t u grad(a )((t)), a=1

m
Ya (q) = grada (q),
(grada )i = Gij
a q j
G
Lemma 31. symmetric product between vector elds satises grada : gradb = grad a : b where symmetric product between functions (Beltrami bracket) is: a : a = da , db a b = Gij i q q j
(t)
(t) = gradVavg ((t)) + G (Fdiss ( (t)))

m
Vavg = V +
a,b=1
ab a : b .
Example: locomotion in the roller racer Example: stabilizing a two-link manipulator via oscillations
1 , 2 (rad)
/2
Two-link damped manipulator with oscillatory control at rst joint. The averaging analysis predicts the behavior. (the gray line is 1 , the black line is 2 ).
Summary averaging theorem for standard form averaging theorem for standard oscillatory form averaging for mechanical systems with oscillatory controls analysis via the averaged potential
time (sec) 1 cos t
20
40
u = 1 +
(i) X1 , X2 describe feasible velocities of racer: X1 forward, X2 shape change (ii) racer has single input X2 (iii) symmetric product X2 : X2 has component along X1 hence, racer moves () forward when subject to zero mean input!
Design of oscillatory controls via approximate inversion Objective: design oscillatory control laws for ACCS stabilization and tracking for systems that are not linearly controllable setup: consider ACCS (Q, , Y, D, Y = {Y1 , . . . , Ym }, Rm ) where Y is an ane map of the velocities dene averaging product A[0,T ] as the map taking a pair of two-time scale vector elds into a time-dependent vector eld by A[0,T ] (V, W )(t, q) = + 1 2T 2 1 2T
T 0 0 T 0 1 0 T 0 0 1 1
V (2 , t, q)d2 :
0 1
W (2 , t, q)d2 d1 W (2 , t, q)d2 d1 .
V (2 , t, q)d2 d1 :
Basis-free restatement of averaging theorem Corollary 32. For R+ , consider governing equations
(t)
Controllability assumption and constructions Controllability Assumption: for all a {1, . . . , m}, Ya : Ya Y
b (i) smooth functions a , a, b {1, . . . , m}, such that, for all a {1, . . . , m} m
1 t (t) = Y (t, (t)) + W , t, (t) ,
(ii) q W (, t, q), for (, t) R+ R+ , are commutative Then, the averaged forced ane connection system is
(t)
(i) W takes values in Y
Ya : Ya =
b=1
b a Yb
(t) = Y (t, (t)) + A[0,T ] (W, W )(t, (t))
(ii) for T R+ and i N, dene i : R R by i (t) = 4i cos T 2i t T
Problem 33 (Inversion Objective). Given any time-dependent vector eld X, compute two vector elds taking values in Y (i) WX,slow is time-dependent (ii) WX,osc is two-time scales, periodic and zero-mean in fast time scale such that WX,slow + A[0,T ] (WX,osc , WX,osc ) = X (1)
(iii) dene the lexicographic ordering as the bijective map 1 lo : (a, b) {1, . . . , m}2 a < b {1, . . . , 2 m(m 1)} given by a1 lo(a, b) = j=1 (n j) + (b a)
Inversion algorithm For an ACCS with Controllability Assumption, assume

m m
Tracking via oscillatory controls Consider ACCS (Q, , Y, D = TQ, Y = {Y1 , . . . , Ym }, Rm ) satisfying Controllability Assumption and span { Ya , Yb : Yc | a, b, c {1, . . . , m}} = TQ
X(t, q) =
a=1
a (t, q)Ya (q) +

b,c=1,b<c
bc (t, q) Yb : Yc (q)
Problem 34 (Vibrational Tracking). given reference ref , nd oscillatory controls such that closed-loop trajectory equals ref up to an error of order Vibrational tracking is achieved by oscillatory state feedback
Then Inversion Objective (1) is solved by

m
WX,slow (t, q) =
a=1
ua X,slow (t, q)Ya (q),
WX,osc (, t, q) =
a=1
ua X,osc (, t, q)Ya (q)
ua X,slow (t, vq )
ua (t) ref
a1
+
b=1
b1+
where
m a ua X,slow (t, q) = (t, q) + b=1 m m
c=b+1
(ubc (t))2 a ref b (q), 4
b1+
i=b+1
( bi (t, q))2 a b (q) 4
ua X,osc (, t, vq ) =
c=1
lo(c,a) ( )
1 uac (t)lo(a,c) ( ) 2 c=a+1 ref
+
b=a+1 a1
1 ab L Ya ab L Yb ab (t, q), 2 1 ai (t, q)lo(a,i) ( ) 2 i=a+1

m
where the ctitious inputs are dened by

m ref (t) ref (t)Y (t, ref (t)) = a=1 m
ua (t)Ya (ref (t))+ ref

b,c=1 b<c
ubc (t) Yb : Yc (ref (t)) ref
ua X,osc (, t, q) =
i=1
lo(i,a) ( )
Example: A second-order nonholonomic integrator Consider x1 = u1 , x2 = u2 , x3 = u1 x2 + u2 x1 ,
Example: A planar vertical takeo and landing (PVTOL) aircraft

x = cos vx sin vz z = sin vx + cos vz = vx vz = g sin + (k1 /m)vx + (1/m)u2
Controllability assumption ok. Design controls to track (xd (t), xd (t), xd (t)): 1 2 3 1 u1 = xd xd xd xd xd cos 3 1 2 2 1 + 2 2 t d u2 = x2 cos xd 1
x1
1 0 1 0 1 10 20 30 40 50
Q = SE(2) : Conguration and velocity space via (x, z, , vx , vz , ). x and z are horizontal and vertical displacement, is roll angle. The angular velocity is and the linear velocities in the body-xed x (respectively z) axis are vx (respectively vz ). u1 is body vertical force minus gravity, u2 is force on the wingtips (with a net horizontal component). ki -components are linear damping force, g is gravity constant. The constant h is the distance from the center of mass to the wingtip, m and J are mass and moment of inertia.

vz + vx = g(cos 1) + (k2 /m)vz + (1/m) = (k3 /J) + (h/J)u2
x2
1 1
10
20
30
40
50
x3
10
20
30
40
50
Oscillatory controls ex. #2: PVTOL model

Controllability assumption ok. Design controls to track (xd (t), z d (t), d (t)):
PVTOL simulations: trajectories and error

1.4
1

0 1
1.2
Error in x Error in z Error in
u2 =
h J 2 f1 sin d + f2 cos d f1 cos d + f2 sin d cos J h

J d h
0 1
Error
0 5 10 15 20 25
2 J k3 u1 = d + d cos h h
10
15
20
25
1 z
0.8
0.6
,
0.5
0.4
where we let c =
k3 d h
and
0.2
sin(2d ) f1 = md + k1 cos2 d + k2 sin2 d xd + x (k1 k2 )z d + mg sin d c cos d , 2 sin(2d ) f2 = md + z (k1 k2 )xd + k1 sin2 d + k2 cos2 d z d + mg(1 cos d ) c sin d . 2
0.5
0
0 5 10
15
20
25
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Trajectory design at = .01.
Tracking errors at t = 10.
Summary averaging theorem for standard form averaging theorem for standard oscillatory form averaging for mechanical systems with oscillatory controls analysis via the averaged potential inversion based on controllability fairly complete solution to stabilization and tracking problems (i) Introduction
Summary
(ii) Modeling of simple mechanical systems (iii) Controllability (iv) Kinematic reductions and motion planning (v) Analysis and design of oscillatory controls (vi) Open problems
Open problems
Modeling (i) variable-rank distributions in nonholonomic mechanics (ii) ane nonholonomic constraints (iii) Riemannian geometry of systems with symmetry (iv) innite-dimensional systems (v) control forces that are not basic (vi) tractable symbolic models for systems with many degrees of freedom Controllability (i) linear controllability of systems with gyroscopic and/or dissipative forces (ii) controllability along relative equilibria (iii) acccessibility from non-zero initial conditions (iv) weaker sucient conditions for controllability
Kinematic reductions and motion planning (i) understanding when the kinematic reduction allows for low-complexity calculation of motion plans for underactuated systems (ii) motion planning with locality constraints (iii) relationship with theory of consistent abstractions (iv) feedback control to stabilize trajectories of the kinematic reductions (v) design of stabilization algorithms based on kinematic reductions
Analysis and design of oscillatory controls (i) series expansions from non-zero initial conditions (ii) motion planning algorithms based on small-amplitude controls (iii) higher-order averaging and inversion + relationship with higher order controllability (iv) analysis of locomotion gaits
References
R. Abraham and J. E. Marsden. Foundations of Mechanics. Addison Wesley, Reading, MA, 2 edition, 1978. ISBN 0-8053-0102-X. A. A. Agrachev and Y. Sachkov. Control Theory from the Geometric Viewpoint, volume 87 of Encyclopaedia of Mathematical Sciences. Springer-Verlag, New YorkHeidelbergBerlin, 2004. ISBN 3-540-21019-9. S. Arimoto. Control Theory of Non-linear Mechanical Systems: A Passivity-Based and Circuit-Theoretic Approach. Number 49 in Oxford Engineering Science Series. Oxford University Press, Walton Street, Oxford, 1996. ISBN 0-19-856291-8. V. I. Arnold. Mathematical Methods of Classical Mechanics. Number 60 in Graduate Texts in Mathematics. Springer-Verlag, New YorkHeidelbergBerlin, 1 edition, 1978. ISBN 0-387-90314-3. Second edition: ?. J. Baillieul. Stable average motions of mechanical systems subject to periodic forcing. In M. J. Enos, editor, Dynamics and Control of Mechanical Systems (Waterloo, Canada), volume 1, pages 123. Fields Institute, Waterloo, Canada, 1993. ISBN 0-821-89200-2. L. M. Bates and J. Z. Sniatycki. Nonholonomic reduction. Reports on Mathematical Physics, 32(1):444452, 1993.
A. M. Bloch. Nonholonomic Mechanics and Control, volume 24 of Interdisciplinary Applied Mathematics. Springer-Verlag, New YorkHeidelbergBerlin, 2003. ISBN 0-387-095535-6. A. M. Bloch and P. E. Crouch. Kinematics and dynamics of nonholonomic control systems on Riemannian manifolds. In Proceedings of the 32nd IEEE Conference on Decision and Control, pages 15, Tucson, AZ, Dec. 1992. A. M. Bloch, M. Reyhanoglu, and N. H. McClamroch. Control and stabilization of nonholonomic dynamic systems. IEEE Transactions on Automatic Control, 37(11): 17461757, 1992. A. M. Bloch, P. S. Krishnaprasad, J. E. Marsden, and R. M. Murray. Nonholonomic mechanical systems with symmetry. Archive for Rational Mechanics and Analysis, 136 (1):2199, 1996. A. M. Bloch, N. E. Leonard, and J. E. Marsden. Controlled Lagrangians and the stabilization of mechanical systems. I. The rst matching theorem. IEEE Transactions on Automatic Control, 45(12):22532270, 2000. A. M. Bloch, D. E. Chang, N. E. Leonard, and J. E. Marsden. Controlled Lagrangians and
the stabilization of mechanical systems. II. Potential shaping. IEEE Transactions on Automatic Control, 46(10):15561571, 2001. B. Bonnard. Controllabilit de syst`mes mcaniques sur les groupes de Lie. SIAM Journal e e e on Control and Optimization, 22(5):711722, 1984. R. W. Brockett. Control theory and analytical mechanics. In C. Martin and R. Hermann, editors, The 1976 Ames Research Center (NASA) Conference on Geometric Control Theory (Moett Field, CA), pages 148. Math Sci Press, Brookline, MA, 1977. ISBN 0-915692-721-X. F. Bullo. Series expansions for the evolution of mechanical control systems. SIAM Journal on Control and Optimization, 40(1):166190, 2001. F. Bullo. Averaging and vibrational control of mechanical systems. SIAM Journal on Control and Optimization, 41(2):542562, 2002. F. Bullo and A. D. Lewis. Low-order controllability and kinematic reductions for ane connection control systems. SIAM Journal on Control and Optimization, 2003. To appear.
F. Bullo and K. M. Lynch. Kinematic controllability and decoupled trajectory planning for underactuated mechanical systems. IEEE Transactions on Robotics and Automation, 17(4):402412, 2001. P. E. Crouch. Geometric structures in systems theory. IEE Proceedings. D. Control Theory and Applications, 128(5):242252, 1981. C. Godbillon. Gomtrie Direntielle et Mchanique Analytique. Collection Mthodes. e e e e e Mathmatique. Hermann, Paris, 1969. e V. Jurdjevic. Geometric Control Theory. Number 51 in Cambridge Studies in Advanced Mathematics. Cambridge University Press, New YorkPort ChesterMelbourneSydney, 1997. ISBN 0-521-49502-4. J. Koiller. Reduction of some classical nonholonomic systems with symmetry. Archive for Rational Mechanics and Analysis, 118(2):113148, 1992. A. D. Lewis and R. M. Murray. Controllability of simple mechanical control systems. SIAM Journal on Control and Optimization, 35(3):766790, 1997. S. Mart nez, J. Corts, and F. Bullo. Analysis and design of oscillatory control systems. e IEEE Transactions on Automatic Control, 48(7):11641177, 2003.
H. Nijmeijer and A. J. van der Schaft. Nonlinear Dynamical Control Systems. Springer-Verlag, New YorkHeidelbergBerlin, 1990. ISBN 0-387-97234-X. R. Ortega, A. Loria, P. J. Nicklasson, and H. Sira-Ramirez. Passivity-Based Control of Euler-Lagrange Systems: Mechanical, Electrical and Electromechanical Applications. Communications and Control Engineering Series. Springer-Verlag, New YorkHeidelbergBerlin, 1998. ISBN 1-85233-016-3. R. Ortega, M. W. Spong, F. Gmez-Estern, and G. Blankenstein. Stabilization of a class o of underactuated mechanical systems via interconnection and damping assignment. IEEE Transactions on Automatic Control, 47(8):12181233, 2002. J. L. Synge. Geodesics in nonholonomic geometry. Mathematische Annalen, 99:738751, 1928. M. Takegaki and S. Arimoto. A new feedback method for dynamic control of manipulators. Transactions of the ASME. Series G. Journal of Dynamic Systems, Measurement, and Control, 103(2):119125, 1981. A. J. van der Schaft. Hamiltonian dynamics with external forces and observations. Mathematical Systems Theory, 15(2):145168, 1981/82. A. J. van der Schaft. Controllability and observability of ane nonlinear Hamiltonian systems. IEEE Transactions on Automatic Control, 27(2):490492, 1982. A. J. van der Schaft. Symmetries, conservation laws, and time reversibility for Hamiltonian systems with external forces. Journal of Mathematical Physics, 24(8):20952101, 1983. A. J. van der Schaft. Controlled invariance for Hamiltonian systems. Mathematical Systems Theory, 18(3):257291, 1985. A. J. van der Schaft. Stabilization of Hamiltonian systems. Nonlinear Analysis. Theory, Methods, and Applications, 10(10):10211035, 1986. A. J. van der Schaft and B. M. Maschke. On the Hamiltonian formulation of nonholonomic mechanical systems. Reports on Mathematical Physics, 34(2):225233, 1994.

10 1 1 92 9625

Transféré par

Informations du document

Description originale:

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

10 1 1 92 9625

Transféré par

Droits d'auteur :

Formats disponibles

Geom.Ctrl.Mech.

Introduction Geometric Control of Mechanical Systems

Thanks to: Thanks to:

Jorge Corts, Kevin Lynch, Sonia Mart e nez, Milo Zefran s

National Science Foundation

autonomous vehicles recongurable, reactive sensing & computation cheap

Geometric modeling of mechanical systems

Formalize the idea of velocity.

Vector elds Assign to each point x M an element of Tx M.

x1 (t) = X 1 (x1 (t), . . . , xn (t))

vector elds { x1 , . . . , xn } on chart domain. X i xi

Any vector eld X is given in coordinates by X = summation convention).

f In coordinates: L X f = X i xi (directional derivative).

One can show that L X L Y f L Y L X f = L [X,Y ] f [X, Y ] = Y i j X i j X Y . xj xj xi

Single rigid body:

(Obody Ospatial ) R3 b1 b2 b3 SO(3).

This is a free mechanical system.

Conguration manifold (contd) Example 2. Planar rigid body: Q = SO(2) R2 S1 R2 .

Conguration manifold (contd) Example 3. Two-link manipulator: Q = SO(2) SO(2) S S .

b2,1 2 b1,2 b1,1 1 s1

Coordinates (1 , 2 ). Coordinates (, x, y). cos sin 0 sin cos 0 0

1 (1 , 2 ) = (R1 , r 1 ) and 2 (1 , 2 ) = (R2 , r 2 ), where cos 1 sin 1 sin 1 R1 = cos 1 0 0 r 1 = r1 R 1 s 1 , r2 =

Conguration manifold (contd) Example 4. Rolling disk: Q = R2 S1 S1 .

(iii) Body angular velocity: t (t)

cos cos cos sin sin

sin cos sin sin cos

Inertia tensor Rigid body B with mass distribution . Mass: (B) =

Inertia tensor about xc : Ic : R R dened by Ic (v) =

Ij,c (j (t)), j (t)

Kinetic energy metric (contd) Example 8. Planar rigid body: 0 I1,c = 0 , 0 0 J

1 (t) = (R1 (t)R1 ) = (0, 0, ), 1

1 (t) = (R1 (t)R1 ) = (0, 0, 1 ), 1

1 (t) = (R1 (t)R1 ) = ( sin , cos , ), 1

2 (t) = (R1 (t)R2 ) = (0, 0, 2 ), 2

1 KE = 1 m(x2 + y 2 ) + 2 Jspin 2 + 1 Jroll 2 , 2 2 m 0 0 0 0 m 0 0 [G] = . 0 0 Jspin 0 0 0 0 Jroll

Here Q = R S S S with coordinates (x, y, , , ).

Glj 1 Glk Gjk , = Gil + k j 2 q q q l

f. The vector eld

Question: What are these functions i ? jk

is the covariant derivative of Y with respect to X.

Theorem 12. Let G be a Riemannian metric on a manifold Q. Then there exists

a unique ane connection

, called the Levi-Civita connection, such that

j, k {1, . . . , n}. in the given

(i) L X (G(Y, Z)) = G(

i , i, j, k {1, . . . , n}, are the Christoel symbols for jk coordinates.

Furthermore, the Christoel symbols of

are i , i, j, k {1, . . . , n}. jk

Return to EulerLagrange equations Had shown that

(i) Covariant derivative of with respect to itself: i j k qi (t) (t) = ( + jk q q ) q i .

(t) (t) = 0 . accn

A covector eld assigns to each point q Q an element of T Q. q

where is the spatial representation of

Sum over all bodies to get total external force F T Q: F = q

The EulerLagrange equations subject to force F are then equivalent to

(t) = G (F )(t, (t))

y = sin . Thus Dq has {X1 (q), X2 (q)} as basis, where X1 = cos

Dene an ane connection

(t) = PD (G (F )(t, (t))).

ua (t)Ya ((t)) + Y (t, (t)),

ua (t)Ya ((t)) + Y (t, (t)).

Natural local representation: