Slides - Libin Liu
Sampling-based Contact-rich Motion Control Libin Liu* KangKang Yin† Michiel van de Panne‡ Tianjia Shao* Weiwei Xu† * Tsinghua University † Microsoft Research Asia ‡ University of British Columbia Motivation 2/38 Motivation 3/38 Motivation Physically-plausible motion reconstruction Input Output 4/38 Motivation Retargeting to new characters Direct tracking Our open-loop control 5/38 Motivation Retargeting to new environments 6/38 Motivation Synthesis of difficult-to-capture motions 7/38 Outline • Motivation • Related Work • Control Construction – Trajectory-based sampling – Practical implementations – Trajectory-free sampling • Results • Discussion 8/38 Related Work • Motor Control and Contact Dynamics [Kawato 99, Jordan and Wolpert 99, Brubaker et al. 09] • Motion Planning [Kavraki et al. 96, LaValle and Kuffner 00, Choi et al. 03, Yamane et al. 04] • Sampling in Animation [Sims 94, Hodgins and Pollard 97, Twigg and James 07, Sok et al. 07, Wang et al. 09, Wampler and Popović 09] 9/38 Outline • Motivation • Related Works • Control Construction – Trajectory-based sampling – Practical implementations – Trajectory-free sampling • Results • Discussion 10/38 Simulation Basics • PD servo k p - kd ~ • Open Dynamics Engine v0.11 11/38 Trajectory-based sampling Sampling [ ] t t + Δt 12/38 Trajectory-based sampling Steps position time t0 t1 t2 t3 t4 t5 13/38 Trajectory-based sampling Steps t += Δt Sample Simulate Calculate cost Select and prune End? No 14/38 Outline • Motivation • Related Works • Control Construction – Trajectory-based sampling – Practical implementations – Trajectory-free sampling • Results • Discussion 15/38 Practical Implementations Sample generation • Sampling window size – Noise level – Joint activeness – Stability of motion dynamics 16/38 Practical Implementations Feedforward torques [ [] ] t t + Δt 17/38 Practical Implementations Sampling time step • Small vs. Large • Uniform vs. Semantic 18/38 Practical Implementations Simulations in Parallel • Master-worker model Master Worker sample selection &maintenance Worker sample generation & simulation & cost evaluation Worker 19/38 Trajectory-based sampling Sample Cost Evaluation E wp E p wr Er we Ee wb Eb • Pose Control : • Root Control: • End-effector Control: • Balance Control: 20/38 Practical Implementations Sample Pruning greedy diversified cost cost f t t x 21/38 Outline • Motivation • Related Works • Control Construction – Trajectory-based sampling – Practical implementations – Trajectory-free sampling • Results • Discussion 22/38 Idling Trajectory-free Sampling 23/38 Idling Dynamic RRT: DoF pruning 24/38 Idling Dynamic RRT: Space Pruning 25/38 Outline • Motivation • Related Works • Control Construction – Trajectory-based sampling – Practical implementations – Trajectory-free sampling • Results • Discussion 26/38 Results Performance Trial walk run sideways roll forward roll backward roll get-up kip-up Duration (s) Reconstruction time (s) 5.2 143 2 51 3 78 3 78 2.1 57 3.5 93 6.6 184 1400 samples for each iteration with success rate > 80% ~25 times slower than real time on cluster of 80 cores 27/38 Results Control Reconstruction 28/38 Results Motion Transformation 29/38 Results: Motion Retargeting 30/38 Results Trajectory-free Control Construction 31/38 Results Motion Composition 32/38 Outline • Motivation • Related Works • Control Construction – Trajectory-based sampling – Practical implementations – Trajectory-free sampling • Results • Discussion 33/38 Advantages of Sampling • Derivative-free • General • Robust wrt. local minima • Easy to parallelize 34/38 Limitations • Robustness of Reconstruction: <100% • Robustness of Control: open-loop • Generalization: trajectory-tracking nature • Smoothness: can be jerky 35/38 Conclusions • Sampling-based reconstruction method general, robust, parallelizable • Contact-rich tasks get-up, rolling, idling • Unified framework inverse dynamics, motion transformation and retargeting 36/38 Acknowledgement • Our mocap subjects • Organizations and authors who generously shared their mocap data • Xiao Liang, Teng Gao, Kevin Loken 37/38 Thank you
© Copyright 2024