Precomputed Search Trees: Planning for Interactive GoalDriven Animation

1
Precomputed Search Trees Planning for
Interactive Goal-Driven Animation

• Manfred Lau and James Kuffner
• Carnegie Mellon University

2
Motion Planning approach
Inputs
Output
3
Behavior Planner
Lau and Kuffner. Behavior Planning for
Character Animation. SCA 05
4
Motivation

• Efficient algorithm for
• large number of characters
• global planning
• re-plan continuously in real-time
• dynamic environment
• complex motions including jump, crawl, duck,
  stop-and-wait

5
Main contribution

• Precomputed Search Tree

6
Traditional Planning

• 50,000 µs for 1 s of motion

7
Precomputed Search Trees

• 250 µs for 1 s of motion

8
Overview
Environment
FSM
9
Overview
Environment

• Precompute

1) Search Tree
FSM
10
Overview
Environment

• Precompute

2) Gridmaps
FSM
11
Overview
Precompute
Runtime
Environment
1) Search Tree
FSM
2) Gridmaps
12
Overview
Precompute
Runtime
Environment
1) Search Tree
1) Map Obstacles
FSM
2) Gridmaps
13
Overview
Precompute
Runtime
Environment
1) Search Tree
1) Map Obstacles
FSM
2) Path Finding
2) Gridmaps
14
Overview Distant Goal
Coarse-Level Planner
15
Overview Distant Goal
Coarse-Level Planner
Repeatedly select sub-goal and run each sub-case
16
Related Work

• Motion Planning
• Kuffner 98Shiller et al. 01Bayazit et al.
  02Choi et al. 03Pettre et al. 03Sung et al.
  05Koga et al. 94Kalisiak and van de Panne
  01Yamane et al. 04

GlobalNavigation
Choi et al. 03
Manipulation andwhole-body motions
17
Related Work

• PrecomputationLee and Lee 04Reitsma and Pollard
  04
• Re-playing original motion capture data
• Arikan and Forsyth 02Kovar et al. 02Lee et al.
  02Pullen and Bregler 02Gleicher et al. 03Lee
  et al. 06

Lee and Lee 04
Kovar et al. 02
18
Related Work

• Motion Vector Fields / Steering Approaches
• Brogan and Hodgins 97Menache 99Reynolds
  99Mizuguchi et al. 01Treuille et al. 06

Treuille et al. 06
19
Advantages of our approach

• Precomputed Search Trees
• many characters re-plan continuously in real-time
• global planning as opposed to local policy
  methods
• complex motions jump, crawl, duck,
  stop-and-wait
• one tree can be used for all characters, and
  different environments

20
Environment Representation
Special regionsfor crawl/jump
Obstacle Growth in Robot Path PlanningUdupa
77Lozano-Pérez and Wesley 83
21
Behavior Finite-State Machine
22
Precompute

• 1) Search Tree

2 levels, 3 behavior states
23
Precompute

• 1) Search Tree

24
Precompute

• 1) Search Treerepresents all states reachable
  from current state

5 levels, 7 behavior states
25
Precompute

• 1) Search Tree Pruned to 10 MB

exhaustive
pruned
26
Precompute
2) Environment Gridmapused to identify the tree
nodes that are blocked by obstacles
27
Precompute
2) Goal Gridmapused to efficiently extract all
paths that reach goal from start state
28
Runtime

• 1) Map obstacles to Environment Gridmap

29
Runtime

• 1) Map obstacles to Environment Gridmap

30
Runtime

• 2) Path Finding reverse path lookup (vs.
  forward search)

31
Runtime

• 2) Path Finding take shortest path that
  reaches goal

32
Runtime

• 2) Path Finding take shortest path that
  reaches goal

33
Runtime

• 2) Path Finding take shortest path that
  reaches goal

34
Motion Generation / Blending

• Sequence of behaviors ? converted to actual
  motion
• Blending at frames near transition points
• Linearly interpolate root positions
• Smooth-in, smooth-out slerp interpolation for
  joint rotations

35
Planning to distant goals

• Only up to specific level

36
Intermediate goal points
Apply precomputed tree repeatedly
37
Intermediate goal points
Apply precomputed tree repeatedly
38
Intermediate goal points
Apply precomputed tree repeatedly
39
Distant goal example
Run coarse bitmap planner first
40
Distant goal example
Find sub-goal
Run sub-case
41
Distant goal example
Find sub-goal
Run sub-case
42
Distant goal example
Final solution
43
Distant goal example
Final solution
44
Result speedup

• Precomputed Trees
  A-search
• Avg. runtime or 3,131
  550,591search time (µs)
  176 times faster
• Avg µs per frame 7.95
  1,445
• Avg pathcost 361
  357
• Avg time of synthesized 13,123,333
  12,700,000motion (µs)
• Real-time speedup 4,191 times
  23 times

45
Tradeoff Motion Quality vs. Memory
exhaustive tree
46
Single Character Mode

• complete solution path for one character
  continuously re-generated, as the user changes
  environment
• large environment (70 by 70 meters), takes 6 ms
  to generate full path

47
Multiple Character Mode

• execute runtime path finding phase only after
  we start rendering the first frame from the
  previous partial path
• precompute blend frames (20 motion clips),
  precompute all pairs
• separate gridmaps for collision avoidance between
  characters
• same precomputed tree for all characters

48
Summary

• Precomputed Search Tree

• Advantages of our approach
• large number of characters
• global planning
• re-plan continuously in real-time
• complex environment
• complex motions

49
Summary

• Precomputed Search Tree

• Advantages of our approach
• large number of characters
• global planning
• re-plan continuously in real-time
• complex environment
• complex motions

Thank you! Questions.