Adapting Simulated Behaviors For New Characters

   

Jessica K. Hodginsand

Nancy S. Pollard

presentation by Barış Aksan

Introduction

To make human-like characters be useful in animations and virtual environments, we must be able to create new, appealing characters easily. Appealing human motion has several components:  The kinematics and dynamics of the figure must be

physically correct

The control algorithms must make the figure perform in ways that appear natural and are stylistically

Dynamic Simulation

Animated motions in the paper are computed using dynamic simulation.

Each simulation contains: The equations of motion for rigid-body

model Constraint equations for the interaction

with the ground Parametrized control algorithms A graphical image A user interface for changing the

parameters

Dynamic Simulation (cont.)

During each simulation time step: Control algorithm computes desired

positions and velocities of joints Proportional-Derivative servos compute

joint torques Equations of motion is integrated with

internal joint torques and external forces and torques from ground or other objects

Graphical Models

Either modeled or purchased Intermediate models for the morph

scenes can be constructed using squares and cylinders

Dynamic Models

Derived from graphical models. Mass and moment of inertia of each body

part is computed using density. There are various joints and certain

controlled degrees of freedom of joints in the models.

Each internal joint has a torque source. Equations of motion were generated

using SD/FAST. Points of contact with ground are

modeled using constraints with Baumgarte stabilization.

THE CONTROLLED DEGREES OF FREEDOM FOR THE DYNAMIC MODEL OF THE RUNNER.

Running Control Algorithms

Running is a cyclic behavior consisting different phases for legs: Flight Heel contact Heel and metatarsus contact Metatarsus contact

Since different phases need different control actions, a state machine is used to select the active control actions.

Running Control Algorithms (cont.) To generate steady-state running;

Forward speed, Flight duration, Balancemust be maintained by the control system.

Those are adjusted in the search step.

Running Control Algorithms (cont.) Foot is positioned at touchdown to

correct errors in forward speed and balance

To reduce disturbances at touchdown, ground speed matching is used.

To control flight duration, ankle and knee joints are extended.

Proportional-Derivative servos are used to compute torquesa and cause the body move to desired roll(0), pitch(slightly forward), yaw(0) values.

Running Control Algorithms (cont.)

k and kv are proportional and derivative gains.

d: desired

For each internal joint, the control equation:

Bicycling Control Algorithms

The goal is controlling: Balance Speed Facing direction

The rider; navigates by applying forces to the

handlebars, controls speed by applying forces to the

pedals.

Scaling

• A control system tuned for one dynamic model, will not work on a different model.

Images showing the result of using the control system designed for the The running motion for a model that

is halfway between the man and the woman.

Geometric Scaling

Assumes; Uniform scaling in all directions, Gravity is same for both characters. Geometric scaling is applied to:

The state of the system Gains for PD joint servos Values and constants used to control

motion (desired values) Integration time step

Geometric Scaling (cont.)

Simple and straightforward Only requires control parameters and the

new model Scaling factor is important

leg length is a good one for running wheel radius is a good one for bicycling

MEASUREMENTS OF MASS AND SIZE OF BODİES AND BICYCLE PARTS

Mass Scaling

Adapts the control system to differences not captured with geometric scaling.

Scales masses and moments of inertia A system with same link lengths has k’

gains,

If link lengths also differ,

Mass Scaling (cont.)

Requires selecting relevant body segments for each gain

Still an approximation because; Depend on a single ratio, Each gain is scaled based on a subset of

body parts Moment of inertia is assumed to be

constant despite changing angles

SCALING RULES

 

Tuning the Motion

This step is a search over high-level parameters.Must result in stable and repetable motion.

Different tunings presented in paper:

• Running• Bicycling• Metamorphosis

Running

The search is restricted to five high-level control parameters:

• Ground speed matching (affects running speed)

• Pitch angle• Timing of thrust (affects duration of flight)• Extension of ankle (affects duration of flight)• Extension of knee (affects duration of flight)

Running (cont.)

The tuning process requires several stages

Intermediate characters are used

An intermediate model that runs approximately 10 seconds with geometric and mass scaling is worth tuning.

Evaluation Function

Values assigned to search parameters define a control system.

Runner is commanded to run for a fixed duration (15 seconds in the paper)

Contains penalties for: Falling Errors in velocity Head acceleration Deviations in roll, pitch, yaw between

strides

Bicycling

The search parameters for tuning bicycling behavior:

Stiffness of arms and shoulders Control of roll and yaw of the bicycle Control of the handlebars and pedals

Different characters may require new bicycle designs.

Metamorphosis

The algorithm can be used for online metamorphosis from one model to another

Equally spaced models between the two models are needed

Has problems because changes in a way that violates physical laws. (change of mass)

Discussion

A hybrid approach is selected

Would the same approach work for new control systems? Diving Vaulting

Is it adaptable to a wide variation of models Toe-strike runners instead of heel-strike runners

Kinematic constraints such as ground or bicycle contact for different link lengths need to be resolved

A COMPARISON OF DATA FROM THE BIOMECHANICAL LITERATURE WITH DATA RECORDED FROM THE SIMULATED RUNNERS.