A Grasp-based Motion Planning Algorithm for Character Animation Maciej Kalisiak and Michiel van de Panne Department of Computer Science, University of Toronto Presented By: Kevin Hufford
Movie [Insert Movie] [Insert Movie] [character climbing through tight passage] [character climbing through tight passage]
Unstructured Env. Navigation Multiple Constraints Multiple Constraints Environment, kinematic, balance, posture Need For Two Decision Types Need For Two Decision Types Discrete Contact: what handholds to use; step on or over obstacles Contact: what handholds to use; step on or over obstacles Continuous Other decisions once contact points are chosen Other decisions once contact points are chosen
Motion Planner Overview
Grasp Point Types Only hands and feet can grasp Only hands and feet can grasp Load-bearing (footholds) Pendent (handholds) Hybrid (either)
Randomized Path Planning (RPP)/Gradient-Descent Gradient Descent Step Gradient Descent Step Configuration change to bring character closer to goal configuration (q target ). Distance-to-goal Metric (C-space potential) Distance-to-goal Metric (C-space potential) P(q) is shortest free-space path (This paper: one control point at COM) Occupancy map of environment distance map Evaluate P(q+ q) for stochastic choices of q Choose q that provides largest collision-free decrease in P
Avoiding Local Minima: Random Walk Random Walk (Brownian Motion) Random Walk (Brownian Motion) At each coordinate j: 1/3
Backtracking (Deep Local Minima) Backtracking (Deep Local Minima) Randomly pick new configuration with whole solution as domain Avoiding Local Minima: Backtracking
Additional Features Locomotion Preferences Locomotion Preferences Smoothing Smoothing Linear interpolation for parts of trajectory with same grasp
Limitations/Future Work Only 2-D so far Only 2-D so far Occasional unstable or gravity-defying postures Occasional unstable or gravity-defying postures Machine learning methods could be applied Machine learning methods could be applied