Fast Synthetic Vision, Memory, and Learning Models for Virtual Humans

1
Fast Synthetic Vision, Memory, and Learning
Models for Virtual Humans
2
Purpose

• Model synthetic vision, memory, and learning
• Quickly synthesize motion from goals

3
Introduction

• Virtual robot
• Combines path planner and controller
• Internal record of perceived objects and states

4
Related Work

• Virtual perception
• Model information flow to character

5
Synthetic Vision

• Determine what is currently visible to character
• Speed ability to handle dynamic environments

6
Synthetic Vision - cont.

• Render unlit model of scene from characters POV
• List of visible objects combined with each
  objects location determines observations

7
A character in a virtual office
True color False Color
8
Internal Representation Memory

• Internal model
• Object geometry from environment and observed
  states

9
Perception-Based Navigation

• Character has set M of observations
• Observations represented as (objIDi, Pi, Ti, vi,
  t)
• M updated at regular intervals

10
Basic sense-plan-control loop (static
environments)
11
Perception-Based Navigation - cont.

• Dynamic environments

12
Perception-Based Navigation - cont.

• Problem Truly missing vs. obscured
• Solution Re-run vision module

13
Revised sense-plan-control loop (dynamic
environments)
14
Learning and Forgetting

• Temporal models
• Different memory rules for different objects
  (logical or deductive model)

15
Experimental Results

• Tested on SGI InfiniteReality2
• Click and drag goals and obstacles

16
1
2
3
4
17
A character exploring unknown mazes
18
Conclusions

• Efficient in storage and update times
• Flexible
• Bottlenecks at synthetic vision model (double
  rendering)