SIMBICON: Simple Biped Locomotion Control

1
SIMBICON Simple Biped Locomotion Control

• KangKang Yin, Kevin Loken, Michiel van de Panne

2
Outline

• Motivation
• Introduction
• Related works
• Balance control strategy
• Mocap-based controllers
• Feedback error learning
• Results
• Conclusions

3
Motivation

• Simple kinematic model
• Cannot predict what happens with an unexpected
  step.
• Cannot respond to being pushed.
• Physically simulated bipeds
• Recover balance realistically from unexpected
  steps and pushes.

Preview
4
Introduction

• Give animated characters good awareness of
  balance for unexpected changes in environment.

5
Related Works

• Data-driven kinematic algorithms
• Mukai and Kuriyama 2005 Kwon and Shin 2005
  Shin and Oh 2006 Hech and Gleicher 2006
• Resequencing and interpolating motion outcomes
• Trajectory optimization
• Fang and Pollard 2004 Liu et al. 2005
  Sulejmanpasic and Popovic 2005 Chai and Hodgins
  2007
• Modeled using optimization criteria incl.
  satisfying user-specified constraints

6
Related Works (cont.)

• Controllers
• Faloutsos et al. 2001 Zordan and Hodgins 202
• physics-based character simulation actually
  models how humans walk.

7
Related Works (cont.)

• Biped locomotion control
• Zero moment point (ZMP) techniques
• Vukobratovic 1969 Honda ASIMO
• Passive dynamics walking
• McGeer 1990 Kuo 1999 Ruina et al. 2001
• Central pattern generators
• Taga et al. 1991
• Reinforcement learning
• Tedrake et al. 2004 Morimoto et al. 2004
• Hopping controllers
• Raibert 1986 Raibert and Hodgins 1991

8
Balance control strategy

• Step 1 develop a cyclical base motion

FSM
Motion capture
9
Finite State Machine
state 0 ? state1 after 0.3s state 1 ? state
2 after swing foot contact
10
Finite State Machine

• Drive each joint to its desired local angle.
• Proportional-derivative (PD) control

Respond to changes.
Damp
11
Balance Control Strategy

• Step 2 control torso and swing-hip wrt world
  frame

12
Balance Control Strategy

• Step 3 COM feedback

v
v
COM velocity matters.


v0
v0


COM position matters.
dlt0
dgt0
13
Balance Control Strategy

• Step 3 COM feedback

Base controller
Continuous feedback
14
Moving to 3D

• Apply the same control ideas to both sagittal and
  coronal planes

15
FSM Design
?t
?t
cd
cv
GUI
4 states
16
Mocap-based Controllers

• Import motions into a dynamic settings.
• Process
• Compute an average gait cycle.
• Base controller tracks the average cycle.
• Apply the same balance strategies.

17
Mocap-based Controllers

• Compute an average gait cycle.
• Fourier analysis to a right hip
• Extract period T of the walking cycle.
• Filters original data to a smooth periodic motion
  T.
• Determine F from T.

Time tm of Largest right hip flex
Foot contact (transition between the states)
State 1, State 3 ignored. F 0.5 / 1 State 0 ?
3 State 3 ? 0
In place of FSM,
18
Mocap-based Controllers

• Base controller tracks the average cycle.
• T serves as a target trajectory.
• Joint angles track using
  PD-controllers.

19
Mocap-based Controllers

• Apply the same balance strategies.
• This time, PD controller is

20
Mocap-based Controllers

• Precision of imitation
• Mismatch of physical parameters
• Tracking rather than anticipating

21
Stiff Response Problems

• Stiff response to perturbations / oscillations
  of the torso pitch angle
• Always reacting to the movement of the hip rather
  than anticipating.
• ? Need to reduce feedback torques.
• See the phenomena from the video.

22
Learning Feed-forward Control
Kawato 1990
23
Feedback Error Learning

• As an inverse model
• Take advantage of being cyclic.
• Learn cyclic motion as a function of phase.

24
Implementation of FEL

• Divide the phase F uniformly into N bins.
• current phase bin
• Blend feed-forward and feed-back torques for each
  bin.

25
Simulation Details 2D Model

• Newton-Euler dynamics formulation
• 70kg trunk, 10kg legs
• Penalty force ground model
• Friction coefficient µ0.65
• Joint limits, torque limits

26
Simulation Details 3D Model

• Open Dynamics Engine
• 28 internal DOF
• Friction coefficient µ0.8
• Joint limits, torque limits

27
Results

• See the video.

28
Results

• Settings

3D model Same as Laszlo et al. 1996 simulated
using Open Dynamics EngineODE PD gain values
kp 300 Nm/rad kd 0.1kp
2D model 7-link planar biped Simulated using
Newton-Euler PD gain values kp 300 Nm/rad
kd 30 Nms/rad
29
Conclusions

• Simple, robust feedback mechanism
• Feedforward strategy that allows for more natural
  low tracking gains
• Real-time, balance-aware physics-based
  characters, with style from mocap