-
1R. Gregg and M. Spong CDC 2009, Shanghai, China 1
Reduction-Based Control of Branched Chains: Application to
3D Bipedal Torso Robots
Robert D. Gregg*Coordinated Science LaboratoryDepartment of
Electrical and Computer EngineeringUniversity of Illinois at
Urbana- Champaign
Mark W. Spong, DeanErik Jonsson School of Engineering and
Computer ScienceDepartment of Electrical Engineering
University of Texas at Dallas
-
2R. Gregg and M. Spong CDC 2009, Shanghai, China 2
Outline
1. Dynamic Walking Background
2. Symmetry and Reduction
3. Symmetries in Mechanical Systems
4. 3D Bipedal Walking
5. Closing Remarks
-
3R. Gregg and M. Spong CDC 2009, Shanghai, China 3
Human Bipedal Locomotion
Humanoid walking is:• dynamic, involving
“controlled falling”• completely 3D,
involving three planes-of-motion
• complex, involving a branching tree structure with many
coupled DOF
-
4R. Gregg and M. Spong CDC 2009, Shanghai, China 4
Planar Compass-Gait Biped
• Many theoretical results consider a serial-chain biped
constrained to the sagittal plane.
• Passive walking gaits (stable limit cycles) down shallow
slopes [McGeer, 1990].
• Mapped to any slope using potential shaping [Spong &
Bullo, 1995].
-
5R. Gregg and M. Spong CDC 2009, Shanghai, China 5
Decomposing Complex Motion• Most 3D bipeds do not
naturally have stable limit cycles for walking.
• Stable limit cycles may exist in the sagittal plane.
• Propose: Exploit symmetries to extend these gaits to 3D with
reduction-based control, separating sagittal, lateral, and axial
control problems.
-
6R. Gregg and M. Spong CDC 2009, Shanghai, China 6
Building Dynamic Gaits
[Gregg and Spong 2008-09]
• Goal: Method to construct straight-ahead and turning gaits for
humanoid bipeds to quickly and efficiently navigate 3D space:
-
7R. Gregg and M. Spong CDC 2009, Shanghai, China 7
Lagrangian Mechanics
• Given config. space , a mechanical system is described by and
Lagrangian function
• Integral curves satisfy E-L equations
n x ninertia/mass matrix
ddt∂L∂q̇ − ∂L∂q = u
L(q, q̇) = K(q, q̇)− V (q)
=1
2q̇TM(q)q̇ − V (q),
(q, q̇) ∈ TQQ
⇐⇒ M(q)q̈ + C(q, q̇)q̇ + ∂∂qV (q) = u
-
8R. Gregg and M. Spong CDC 2009, Shanghai, China 8
Example of Symmetry
• Letting , variable is cyclic if
• Lagrangian invariance under rotations of .• Generalized
momentum
• Uncontrolled E-L equations show
• Conservation law: is constant.
0ddt
∂L∂q̇1− ∂L∂q1 =
ddtp1 = 0.
p1 = J1(q, q̇) :=∂L∂q̇1
p1
Q = Tn q1
S1 q1
∂L∂q1
= 0.
-
9R. Gregg and M. Spong CDC 2009, Shanghai, China 9
Symmetry-Based Reduction
• In Routhian reduction, a system with config. space has cyclic
variables :
TQ
mod G
: phase space
momentum map surface
reduced phase space
J−1(μ) :
qi ∈ GiQ = G× S
pi = Ji(q, q̇) = μi
TS :
∂L∂qi
= 0
-
10R. Gregg and M. Spong CDC 2009, Shanghai, China 10
Divided Variables?
Likely to be unstable, e.g., yaw and lean for a biped…
-
11R. Gregg and M. Spong CDC 2009, Shanghai, China 11
Symmetry-Breaking Geometric Reduction
Stabilized cyclic coordinates
Controlled Reduction
Stabilized reduced subsystem
Ji(q, q̇) = λi(qi)Ji(q, q̇) 6= μi
Introduced in [Ames, Gregg, Wendel, Sastry, 2006]
-
12R. Gregg and M. Spong CDC 2009, Shanghai, China 12
Lagrangian Shaping• Lagrangian L with scalar cyclic and vector
:
• Desire closed-loop system corresponding to an almost-cyclic
Lagrangian:
where are special energy shaping terms based on momentum
function
Kaugλ , Vaugλ
λ(q1).
q1
L(q2, q̇) = K(q2, q̇)− V (q2)
=1
2q̇TM(q2)q̇ − V (q2)
Lλ(q, q̇) = K(q2, q̇)− V (q2)+ Kaugλ (q, q̇2)− V augλ (q),
q2
-
13R. Gregg and M. Spong CDC 2009, Shanghai, China 13
Reduced Lagrangian
Given almost-cyclic Lagrangian
The Routhian (reduced Lagrangian):
J(q, q̇) = λ(q1)
Lred(q2, q̇2) = [Lλ(q, q̇)− λ(q1)q̇1]|J(q,q̇)=λ(q1)=
1
2q̇T2M2(q2)q̇2 − V (q2)
Lλ(q, q̇) =12
¡q̇1 q̇
T2
¢µ m1(q2) ??T M2(q2)
¶µq̇1q̇2
¶− V (q2)
+Kaugλ (q, q̇2)− Vaugλ (q)
-
14R. Gregg and M. Spong CDC 2009, Shanghai, China 14
Reduction Revisited
TQ: phase spacereduced phase spaceTS :
= λ(q1)
p1 = J1(q, q̇)
mod G
Lλ(q, q̇)
J−1(λ)
Lred(q2, q̇2)
(q1(t), q2(t), q̇1(t), q̇2(t)) ←→ (q2(t), q̇2(t))
-
15R. Gregg and M. Spong CDC 2009, Shanghai, China 15
Example: An matrix M is recursively cyclic in the first 3
coordinates if it has the form:
=
⎛⎝ m1(qn2 ) m12(qn2 ) M13(qn2 )m12(qn2 ) m2(qn3 ) M23(qn3
)MT13(q
n2 ) M
T23(q
n3 ) M3(q
n4 )
⎞⎠
Finding Symmetries
n× n
where qnj = (qj , qj+1, . . . , qn)T
(q2, . . . , qn)(q3, . . . , qn)(q4, . . . , qn)
M(q2, . . . , qn) =
µm1(qn2 ) M12(q
n2 )
MT12(qn2 ) M2(q
n3 )
¶
-
16R. Gregg and M. Spong CDC 2009, Shanghai, China 16
Extensive Symmetries
General case: An matrix M is recursively cyclic in k coordinates
if it has the form:
n× n
for , where and qnn+1 = ∅.qnj = (qj , qj+1, . . . , qn)T
Cannot use one stage of reduction for all k variables!
Controlled reduction by stages.
M(q2, . . . , qn) =
⎛⎜⎜⎜⎝m1(qn2 ) –— M12(q
n2 ) –––
| . . ....
MT12(qn2 ) mk−1(q
nk ) Mk−1,k(q
nk )
| · · · MTk−1,k(qnk ) Mk(qnk+1)
⎞⎟⎟⎟⎠
k ≤ n
-
17R. Gregg and M. Spong CDC 2009, Shanghai, China 17
• For a branched chain, this holds for some .
• Related to kinetic energy invariance under rotations of
inertial frame [Spong & Bullo, 1995].
Property of Mechanical Systems
z
xy
[Gregg and Spong, 2008]
Theorem: The inertia matrix of any n-DOF serialkinematic chain
is recursively cyclic in n coordinates.
SO(3)
k ≤ n
-
18R. Gregg and M. Spong CDC 2009, Shanghai, China 18
• A hybrid control system has the form
• A hybrid system has no explicit input (e.g., closed-loop
systems):
Hybrid Systems
H
HC :½ẋ = f(x) + g(x)u x ∈ D\Gx+ = ∆(x−) x− ∈ G
x∈ G ∆(x)
D
ẋ = f(x)
D = {x|h(x) ≥ 0} ⊆ TQ
x =
µqq̇
¶ P : G→ Gxj+2 = P 2(xj)δP 2 about x∗
-
19R. Gregg and M. Spong CDC 2009, Shanghai, China 19
Branched Biped Model
x00: θs
y0: ϕ
ZYX-Euler joint at the foot/ankle:
Yaw aboutRoll aboutPitch about
5-DOF config ,
with sagittal configθ = (θs, θt, θns)
T .
q = (ψ,ϕ, θT )T
z: ψ
-
20R. Gregg and M. Spong CDC 2009, Shanghai, China 20
Controlled Reduction by Stages
5-DOF 3D biped(no dynamic gaits)
5-DOF 3D biped with dynamic gaits
4-DOF 3D biped with dynamic gaits
3-DOF planar biped with dynamic gaits
energy shaping
Yaw
Lean
p1 = λ1(ψ) = −α1(ψ − ψ̄)
p2 = λ2(ϕ) = −α2ϕ
-
21R. Gregg and M. Spong CDC 2009, Shanghai, China 21
Straight-Ahead Gait
LES 2-periodic limit cycle along heading :ψ̄
x∗st = P 2st(x∗st)
-
22R. Gregg and M. Spong CDC 2009, Shanghai, China 22
Constant-Curvature Steering
• Constant steering angle induces LES periodic turning gaits
modulo heading change:
s = ∆ψ̄
Hipless 4-DOF Hipless 5-DOF
-
23R. Gregg and M. Spong CDC 2009, Shanghai, China 23
CW-Turning Gait
LES 2-periodic limit cycle (mod yaw) for :s = π/14
modψ(x∗tu(s), s) = P 2tu(s)(x
∗tu(s))
-
24R. Gregg and M. Spong CDC 2009, Shanghai, China 24
Path Planning by Switching
Constant-curvaturewalking arcs
[Submitted ICRA10]
• A 3D biped is a discrete-time switched system where the
switching signal from step-to- step determines the walking
path:
σ(k)
xk+2 = P2σ(k)(xk)
-
25R. Gregg and M. Spong CDC 2009, Shanghai, China 25
Planned Walking Path
-
26R. Gregg and M. Spong CDC 2009, Shanghai, China 26
Locomotor Energeticscet = E/(mgd)
-
27R. Gregg and M. Spong CDC 2009, Shanghai, China 27
谢谢
(Thank You)!
Special thanks to Mark Spong, Tim Bretl, and Jessy Grizzle
Send comments to [email protected]
-
28R. Gregg and M. Spong CDC 2009, Shanghai, China 28
Select Publications•“Asymptotically Stable Gait Primitives for
Planning Dynamic Bipedal Locomotion in Three Dimensions.” Gregg,
Bretl, and Spong. Submitted to 2010 ICRA, Anchorage, AK.
•“Bringing the Compass-Gait Bipedal Walker to Three Dimensions.”
Gregg and Spong. In 2009 IROS, St. Louis, MO.
•“Reduction-Based Control of 3D Bipedal Walking Robots.” Gregg
and Spong. Int. J. of Robotics Research, Pre-print, Accepted
2009.
•“Reduction-based Control with Application to 3D Bipedal Walking
Robots.” Gregg and Spong. In 2008 ACC, Seattle, WA.
•“A Geometric Approach to 3D Hipped Bipedal Robotic Walking.”
Ames, Gregg, and Spong. In 2007 CDC, New Orleans, LA.
•“On the Geometric Reduction of Controlled 3-D Bipedal Walking
Robots.” Ames, Gregg, Wendel, and Sastry. In the 2006 Workshop on
Lagrangian and Hamiltonian Methods for Nonlinear Control, Nagoya,
Japan.
-
29R. Gregg and M. Spong CDC 2009, Shanghai, China 29
Reduction-Based Control
• Energy-shaping control
yields dynamics for controlled reduction.
• Does not cancel natural nonlinearities, only adds shaping
terms.
• Recursively cyclic M allows multiple stages of controlled
reduction!
u = C(q2, q̇)q̇ +N(q2)
+M(q2)Mλ(q)−1 (−Cλ(q, q̇)q̇ −Nλ(q) + v)
-
30R. Gregg and M. Spong CDC 2009, Shanghai, China 30
Mapping Branched Chains• Map branched chain to higher-
order redundant serial chain.
• Wrap around each radiating branch using redundant paths.
• Zero-mass redundant links, constrained redundant joints:
• Constrained redundant system (DAE) projects onto n-DOF
dynamics of branched chain.
ir2 = π, ir6 = π, i
r4 = i5, i
r5 = i6
-
31R. Gregg and M. Spong CDC 2009, Shanghai, China 31
Branched Chain Symmetries
• An n-DOF branched chain is recursively cyclic down to the
m-DOF submatrix, , of the irreducible tree structure.
• Proposition: The irreducible structure starts at the 2nd joint
of the first non-simple branch.
reducible (gray) irreducible
1 ≤ m ≤ n
M(qn2 ) =
⎛⎜⎝ m1(qn2 ) M12(q
n2 )
. . ....
MT12(qn2 ) . . . M5(q
n5 )
⎞⎟⎠M5(q
n5 )
Reduction-Based Control of Branched Chains: Application to 3D
Bipedal Torso RobotsOutlineHuman Bipedal LocomotionPlanar
Compass-Gait BipedDecomposing Complex MotionBuilding Dynamic
GaitsLagrangian MechanicsExample of SymmetrySymmetry-Based
ReductionDivided Variables?Controlled ReductionLagrangian
ShapingReduced LagrangianReduction RevisitedFinding
SymmetriesExtensive SymmetriesProperty of Mechanical SystemsHybrid
SystemsBranched Biped ModelControlled Reduction by
StagesStraight-Ahead�GaitConstant-Curvature
SteeringCW-Turning�GaitPath Planning by SwitchingPlanned Walking
PathLocomotor Energetics谢谢 (Thank You)!Select
PublicationsReduction-Based ControlMapping Branched ChainsBranched
Chain Symmetries