TEN YEARS IN LOCOMOTION CONTROL RESEARCH Jehee Lee Seoul National University
[SIGGRAPH 2010] Lee et al, Data-driven biped control
[SIGGRAPH 2010] Lee et al, Data-driven biped control
[SIGGRAPH 2010] Lee et al, Data-driven biped control
Hubo
Before 2007 After 2007 Simplified dynamics Model Fullbody dynamics Feedback only (stereotyped robotic walking) Feedback and feedforward (motion capture references) Analytic balance strategy Learning from experience Derivative-based optimization (conjugate-gradient, Newton, BFGS, ) Derivative-free optimization (CMA-ES)
Before 2007 After 2007 Simplified dynamics Model (inverted pendulum) Fullbody dynamics Feedback only (stereotyped robotic walking) Feedback and feedforward (motion capture references) Analytic balance strategy Computational model of balancing (regression, learning from experience, optimization at runtime) Derivative-based optimization (conjugate-gradient, Newton, BFGS, ) Derivative-free optimization (CMA-ES)
[SIGGRAPH 2007] Sok et al, Simulating Biped Behaviors from Human Motion Data
[SIGGRAPH 2010] Lee et al, Data-driven biped control
Before 2007 After 2007 Simplified dynamics Model (inverted pendulum) Fullbody dynamics Feedback only (stereotyped robotic walking) Feedback and feedforward (motion capture references) Analytic balance strategy Computational model of balancing (regression, learning from experience, optimization at runtime) Derivative-based optimization (conjugate-gradient, Newton, BFGS, ) Derivative-free optimization (CMA-ES)
Plausibility of Simulation Physical Correctness The simulation is correct with respect to Newton s law of motion No fictional force applies to the body Admissible Control Control force/torques are valid within muscle capacity GRF (ground reaction force) consistent with control force/torque
Plausibility of Simulation Type I (Strongly admissible) Simulation is physically correct and control is admissible Type II (weakly admissible) Simulation is physically correct, but control may not be admissible ex) GRFs are computed as optimization parameters independent of joint torques Type III (Visually plausible) Physical correctness is not guaranteed ex) Fictional force may apply at contact points
Dynamics Energertic Stability Balance Agility Low-energy Muscle Skin Tendon Modeling Pertubation Static Skeleton Applications Gait Analysis Humanoid Robot Robustness Video Games Biological Motion Simulation Social Group Quadruped Biped Emotion Fatigue Aging Adapation High-Level Behavior Interaction Flying Type
Dynamics Energertic Stability Balance Agility Low-energy Muscle Skin Tendon Modeling Pertubation Static Skeleton Applications Gait Analysis Humanoid Robot Robustness Video Games Biological Motion Simulation Social Group Quadruped Biped Emotion Fatigue Aging Adapation High-Level Behavior Interaction Flying Type
L Gait2562 (25 DOFs, 62 muscles) Gait2592 (25 DOFs, 92 muscles) Fullbody (39 DOFs, 120 muscles) [SIGGRAPH Asia 2014] Lee et al, Many-Muscle Humanoids
[SIGGRAPH Asia 2014] Lee et al, Many-Muscle Humanoids 18
Dynamics Energertic Stability Balance Agility Low-energy Muscle Skin Tendon Modeling Pertubation Static Skeleton Applications Gait Analysis Humanoid Robot Robustness Video Games Biological Motion Simulation Social Group Quadruped Biped Emotion Fatigue Aging Adapation High-Level Behavior Interaction Flying Type
Unilateral Painful Ankle Plantar Flexor Patients tend to reduce the use of the ankle plantar flexors
Painful Joints on Unilateral Limb Patients tend to reduce contact force
Painful Left Ankle Plantar Flexor Painful Joints on Left Leg
Waddling Gait Bilateral Gluteus Medius & Minimus Weakness Upper body swing laterally
Trendelenburg Gait Unilateral Gluteus Medius & Minimus Weakness
Dynamics Energertic Stability Balance Agility Low-energy Muscle Skin Tendon Modeling Pertubation Static Skeleton Applications Gait Analysis Humanoid Robot Robustness Video Games Biological Motion Simulation Social Group Quadruped Biped Emotion Fatigue Aging Adapation High-Level Behavior Interaction Flying Type
Balance and Stability Under what conditions is human gait more stable? What factors affect the level of stability? Are simulated walking as stable as human walking? Do the factors that affect human gait also influence controller stability?
[SIGGRAPH Asia 2015] Lee et al, Push-Recovery Stability
Four factors that affect gait stability Level of crouch Walking speed Magnitude of push Timing of push Crouch Gait is more stable than Normal Gait It detours less if it walks faster, push is weaker, and push happens later in the swing phase Similar trends for humans and simulation
Applications in Clinical Gait Analysis Surgery improves cerebral palsy gait by lengthening tight muscles/tendons and fixing bone deformity Predictive simulation of post-operative gaits from pre-operative motion capture and surgery planning
Dynamics Energertic Stability Balance Agility Low-energy Muscle Skin Tendon Modeling Pertubation Static Skeleton Applications Gait Analysis Humanoid Robot Robustness Video Games Biological Motion Simulation Social Group Quadruped Biped Emotion Fatigue Aging Adapation High-Level Behavior Interaction Flying Type
Papers & Videos are available at http://mrl.snu.ac.kr/