Yoonsang LeeSungeun Kim
Jehee Lee
Seoul National University
Data-Driven Biped Control
Biped Control
?Human Biped character
Biped Control is Difficult
• Balance, Robustness, Looking natural• Various stylistic gaits
PETMANBoston Dynamics
ASIMOHonda
HUBOKAIST
Issues in Biped Control
Naturalness
Robustness
Richness
Interactivity
human-like natural result
maintaining balance
variety of motor skills
interactive control via user interface
Goal
As realistic as motion capture data
Robust under various conditions
Equipped with a variety of motor skills
Controlled interactively
Naturalness
Robustness
Richness
Interactivity
Related Work• Manually designed controller
– [Hodgins et al. 1995] [Yin et al. 2007]
• Non-linear optimization– [Sok 2007] [da Silva 2008] [Yin 2008] [Muico 2009] [Wang 2009] [Lasa 2010] [Wang 2010] [Wu
2010]
• Advanced control methodologies– [da Silva 2008] [Muico 2009] [Ye 2010] [Coros 2010] [Mordatch 2010]
• Data-driven approach– [Sok 2007] [da Silva 2008] [Muico 2009] [Tsai 2010] [Ye 2010] [Liu 2010]
Our Approach
• Control methods have been main focus– Machine learning, optimization, LQR/NQR
• We focus on reference data– Tracking control while modulating reference data
Our Approach
• Modulation of reference data
• Balancing behavior of human
• Importance of ground contact timings
Advantages
• Do not require– Non-linear optimization solver– Derivatives of equations of motion – Optimal control– Precomputation
Easy to implement & Computationally efficient
Advantages
• Reference trajectory generated on-the-fly can be used
Any existing data-driven techniques can be used to actuate physically simulated bipeds
Overview
forward dynamics simulation
animation engine
user interaction
data-driven control tracking control
Overview
forward dynamics simulation
user interaction
data-driven control tracking control
animation engine
• High-level control through user interfaces• Generate a stream of movement patterns
Animation Engine
motion fragments
query
motion DBpattern generator
user interaction
stream of movement patterns
• Collection of half-cycle motion fragments
• Maintain fragments in a directed graph
Motion Database
motion capture data motion fragments
Overview
forward dynamics simulation
user interaction
tracking control
animation engine
data-driven control
Data-Driven Control
• Continuous modulation of reference motion
• Spatial deviation– SIMBICON-style feedback balance control
• Temporal deviation– Synchronization reference to simulation
Balancing
...reference motion
simulation
frame n frame n+1 frame n+2
...
...
frame n frame n+1 frame n+2
Balancing
target pose
...reference motion
simulation
...
...
frame n frame n+1 frame n+2
Balancing
tracking
...reference motion
simulation
...
...
frame n+1 frame n+2frame n
Balancing
tracking
...reference motion
simulation
...
...
Balance Feedback
• Near-passive knees in human walking
• Three-step feedback– stance hip– swing hip & stance ankle– swing foot height
Balance Feedback
• Biped is leaning backward
?
reference motionat current frame
reference motionat next frame
simulation
• Stance Hip
Balance Feedback
target poseat next frame
reference frame simulation
• Swing Hip & Stance Ankle
Balance Feedback
target poseat next frame
reference frame simulation
Balance Feedback
• Swing Foot Height
target poseat next frame
reference frame simulation
Feedback Equations
Stance hip
Swing hip
Stance ankle
Swing foot height
reference frametarget pose
Feedback Equations
desired states current states
Stance hip
Swing hip
Stance ankle
Swing foot height
Feedback Equations
parameterstransition function
Stance hip
Swing hip
Stance ankle
Swing foot height
Synchronization
referencemotion
swing foot contacts the ground
Synchronization
current time
referencemotion
simulation
Early Landing
referencemotion
contact occurs!simulation
Early Landing
referencemotion
simulation
dequed
Early Landing
referencemotion
simulation
Early Landing
referencemotion
simulation
warped
Early Landing
referencemotion
simulation
Delayed Landing
referencemotion
not contact yet!simulation
Delayed Landing
referencemotion
simulation
expand byintegration
Delayed Landing
referencemotion
simulation contact occurs!
expand byintegration
Delayed Landing
referencemotion
simulation
warped
Delayed Landing
referencemotion
simulation
Overview
forward dynamics simulation
user interaction
animation engine
data-driven control tracking control
• Compute torques that attempts to follow reference trajectory (ex. PD control)
• We use floating-base hybrid inverse dynamics
Tracking Control
inverse dynamicsdesiredjoint accelerations joint torques
external forces
Why does this simple approach work?
• Human locomotion is inherently robust
• Mimicking human behavior– Distinctive gait serves as a reference trajectory– We do modulate the reference trajectory
Discussion
• We do not need optimization, optimal control, machine learning, or any precomputation
• Physically feasible reference motion data
• Future work– Wider spectrum of human motions
Acknowledgements
• Thank– All the members of SNU Movement Research
Laboratory– Anonymous reviewers
• Support– MKE & MCST of Korea
Data-Driven Biped ControlYoonsang Lee, Sungeun Kim, Jehee Lee