Download - DARPA Robotics Challenge cga/dw. Day 1
DARPA Robotics Challengewww.cs.cmu.edu/~cga/dw
Day 1
Real TimeFinals
Optimization All The Way Down
• Multi-level optimization:– Footstep Optimization (Discrete + Continuous) – Trajectory Optimization (Continuous)– Optimization-Based Inverse Dynamics: Greedy
continuous optimization (Quadratic Program = QP) for full body at the current instant.
Trials
Multi-Level Control• Trajectory: External forces at contacts drive
center of mass (COM).
Rotation is more complicated:– Constant angular momentum– Rigid body equivalent– General case
• QP: Redundancies and constraints resolved for full body behavior.
xmF
IIFr )(
0)( Fr
dtIdLFr /)()(
Trials
LIPM Trajectory Optimization
X vs, Time
Y vs. Time
COM
Footstep
Footstep
Footstep
COM
COM
X vs. Y
meters
Trials
Optimization-Based “Inverse Dynamics” (QP)
Objectives:• Dynamics• Task Objectives• COM Acceleration• Torque About COM• Reference Pose Tracking• Minimize Controls
Constraints:• Center of Pressure• Friction Cone• Joint Torque Limits
dq
C
bw
bw
bw
bw
q
Aw
Aw
Aw
Aw
NNNN
33
22
11
33
22
11
M. de Lasa, I. Mordatch, and A. Hertzmann, “Feature-Based Locomotion Controllers,” ACM Transactions on Graphics, vol. 29, 2010.
Stephens
Trials
Stuck on the door
Finals
Handling modeling error and external forces
Finals
Finals
Failed manipulation
Finals
Fall Predictors
Finals
A bad stepFinals
Why Dynamics Matters
Wheels win?• All wheeled/tracked robots plowed through
debris.• All other robots walked over rough terrain.• KAIST – walked on stairs• Nimbro, RoboSimian – no stairs• Leg/wheel hybrids good if there is a flat
floor somewhere under the pile of debris.• Wheeled/tracked vehicles fell: need to
consider dynamics, need to be able to get up (CHIMP, NimbRo), and get un-stuck.
Finals
Whole-Body Locomotion
• No robot used railings, walls, door frames, or objects in the environment for physical guidance, stabilization, or support.
• Some robots used arms to get out of the car.
Finals
Can Atlas Be Human Like?www.cs.cmu.edu/~cga/dw
• Slow (0.5m/s)• Slow (0.6s/step)• Short steps 0.35m
(half human 0.7m)• Flat foot walking• Bent knees• No heel strike or
toe push off• Fixed pelvis
orientation
• Minimal vertical COM acceleration -> GRF
• Minimal horizontal COM acc. -> GRF
• COP fixed at center of foot
• Not robust to external perturbations or modeling error.
• Soft ground, sand,…?
Baby Steps To Human-Like Walking
Current
Current