Dynamic Controllers in Character Animation

Dynamic Controllers in Character Animation

Jack Wang

2

Overview

• Definition• Related Work• Composable Controllers Framework

(2001)• Results• Future Work

3

Physics-based Animation

• Dynamic Controllers vs. Simulation• Active characters.

• Dynamic Controllers vs. Optimal Trajectory• Physics constraints are enforced to a

different degree.• Animator constraints are enforced to a

different degree.

4

Picture

5

Example Controllers

• Open loop, no feed back, does not depend on current state.

• Close loop, output is a function of the current state.• torque = K_s(q – q_des) – K_dq’ is a

Proportional Derivative (PD) controller• Neural-Network type controller (including

Sensor-Actuator Networks)• Pose controller

6

Pose Controller

7

Related Fields

• Biomechanics• Highly detailed models, geared towards

applications in medicine.• Anderson and Pandy, 2001

• Solves for the muscle activation history of half a walk cycle.

• Minimizes metabolic energy per unit distance traveled.

• 810 dimensional optimization problem.

• Etc…

8

Related Fields

• Motor Neuroscience• Interested in how the brain issues motor

commands.• Harris and Wolpert, 1998

• Studied goal-directed arm and eye movements.

• Control signals are corrupted by multiplicative noise.

• Proposes people minimize variance in final position when planning trajectory.

• Etc…

9

Related Fields

• Robotics• Significant overlap with controller-based

animation in terms of interests, especially in humanoid and animal-like robots.

• Large amount of work done in locomotion controllers.

• State of the art Honda robot still doesn’t walk like humans do.

10

Controllers in Animation

• Hand tuned controllers• Human athletics (running, vaulting,

cycling) – Hodgins et al., 1995• Human diving – Wooten and Hodgins,

1996• 3D bipedal walk – Laszlo et al., 1996

11

Controllers in Animation

• Optimization• Simple planar figures - van de Panne and

Fiume, 1993• Virtual creatures – Sims, 1994• Aquatic animals - Grzeszczuk and

Terzopoulos, 1995

12

Controllers in Animation

• Motion capture modification • Tracking upper body movements –

Zordan and Hodgins, 1999

• Interactive animation• Driving planar characters with mouse

input – Laszlo et al., 2000

13

Synthesis Methods

• Design by hand• Optimization

• Could design an initial controller by hand.

• Gradient usually unavailable.

• Reinforcement Learning• Algorithms have been developed to

perform optimization in stochastic environments.

14

Composable Controllers

• People have been synthesizing controllers to perform specific tasks.

• A framework to combine controllers so that more complex tasks can be performed, Faloutsos et al., 2001

15

Controller Abstraction

• Pre-Conditions• Just like in programming, regions in the state-

space that the controller can operate.• Unlike in programming, success is not

guaranteed.

• Post-Conditions• Defines what “success” means.

• Expected Performance• Regions in state-space that are likely for the

controller to succeed, once execution has started.

16

Example (Falling)

17

Supervising Controller

18

Typical Transitions

19

More on Pre-Conditions

• Can be hard to set manually.• Can be formulated as a classification

problem: given initial state and controller, classify success and failure.

• Use Support Vector Machines (SVM)

20

Linear Support Vector Machines• Solves for a hyperplane in state-space that

maximizes the distance to the closest data points (support vectors), constrained by the separation of data.

21

Nonlinear Boundary

• Map data points to higher, possibly infinite dimensional space, where they could be separated by a linear boundary.

• Introduce kernel functions.• Basically, they are inner products of

data in a higher dimensional space.

22

Example

23

Results

24

Observations

• Very robotic motion.• No locomotion capabilities to the 3D

character.• Best results come from highly

dynamic plunging, falling motion.• Controllers cannot be easily adapted

to new models.

25

Future Work

• Need controllers with human gaits.• Need model-independent algorithms

to synthesize controllers.• Need robust controllers. • Already 4 years into the future.

26

Possible Directions

• Synthesize controllers under motor noise• Necessary for robotics, more robust

controllers.• Surprisingly good effect on gait.

(Lawrence et al., 2003)

• Passive Dynamics• Simplify control by engineering models

that can naturally perform unstable tasks. (Collins et al., 2005)

27

(Partial) References• Composable Controllers for Physics-based Character

Animation, Faloutsos et al., SIGGRAPH 2001.• Composable Controllers for Physics-based Character

Animation, Petros Faloutsos, Phd. Thesis, University of Toronto.

• Dynamic Optimization of Human Walking, Anderson and Pandy, JBE, 2001.

• Signal-dependent noise determines motor planning, Harris and Wolpert, Nature, 1998

• Efficient Gradient Estimation for Motor Control Learning, Lawrence et al., UAI, 2003

• Efficient Bipedal Robots Based on Passive-Dynamic Walkers, Collins et al., Nature, 2005

• A Tutorial on Supper Vector Machines for Pattern Recognition. DMKD, 1998