Online Control of Simulated Humanoids Using Particle Belief

Online Control of Simulated Humanoids
Using Particle Belief Propagation
Motivation
• Control simulated humanoid
• Various movements, environment
• Without any pre-computation, motion capture data
• At real time
Simulation Model
: State ( pose and velocity )
: Control. ( Desire joint angle )
Character model
( 15 bones, 30 DOF )
Objective
( for balancing )
vel : speed of COM
com : horizontal distance of COM from the feet
y
: y position of COM relative to feet
feet : distance between each foot
w
: angular speed of the pelvis
up : difference between the pelvis up vector and global up vector
fwd : head facing direction
damage : 10000 if the character’s head touches the environment
Previous work
Reference
Motion
Simulation
Fall down
Previous work
Reference
Motion
Change reference motion
Previous work
Reference
Motion
Simulation
Initial Pose
Result
Control
Simulation
Initial Pose
Result
Control
Simulation
Pick best sample
Result
Control
Simulation
Result
Control
Simulation
Result
Control
Simulation
Result
Control
Simulation
N : # of samples ( = 32 )
K : Planning horizon ( = 1.2s, 36 time step )
Simulation Model
: State ( pose and velocity )
: Control. ( Desire joint angle )
Character model
( 15 bones, 30 DOF )
Sampling
Resampling
Backwards local refinement
Using previous
trajectories as a prior
Probability Model
Probability Model
Probability Model
Control as Markov Random Field
Belief Propagation
Particle Belief Propagation
Particle Belief Propagation
Particle Belief Propagation
Result
Control
Simulation
Resampling
Local Refinement
Operation Over Multiple Frames
Previous
Step
Current
Step
Sampling
Operation Over Multiple Frames
Previous
Step
Current
Step
Total Algorithm