automated human motion in constrained environments maciej kalisiak [email protected]
TRANSCRIPT
Automated human motion in constrained environments
Maciej [email protected]
Introduction
• human character animation• constrained environments• kinematic method• currently 2D, extendible • sample solution
Path Planning
• piano mover’s problem
• given: start and goal configurations
• find connecting path
Application to Human Motion
Approach
• starting point: RPP
• additions:– moving while in contact with environment– notion of comfort– knowledge of human gaits
Understanding RPP
• Randomized Path Planning
• a path planning algorithm
Simplest “Planner”
• character’s state: q
• repeated perturbations,i.e., Brownian motion
• repeat until goal reached
• discretize into grid
• potential = Manhattan distance to goal
• flood-fill
Building a Potential Field
Gradient Descent
• character point mass• sample q’s neighbourhood• pick sample with largest
drop in potential• iterate until goal reached• not feasible analytically
Local Minima
• gradient descent stops at any minimum
• use random walks to escape– Brownian motion of predetermined duration
• use backtracking if minimum too deep– revert to a previous point in solution,
followed by a random walk
Deep Minimum Example
Smoothing
• solution embodies complete history of search process
• also very noisy
• a trajectory filter post-process is applied– removes extraneous motion segments– makes remaining motion more fluid
Modifications
• grasps and grasp invariants
• comfort heuristic system
• gait finite state machine
• grasp-aware gradient descent, random walk, smoothing filters
Character Structure
• 10 links
• 9 joints
• 12 DOFs
• frequent re-rooting
Grasp Points
• represent potential points of contact
• three types
• reduce the grasp search space
• summarize surface characteristics
Grasp Invariants
• each gait dictates:– the number of grasps– the types of grasps
• enforced by the GFSM
• rest of planner must not alter existing grasps
Motion without Heuristics
Heuristic System
• each heuristic measures some quality of q
• D(q): overall discomfort, a potential field
• getting comfy: gradient descent through D(q)
Implemented Heuristics
The Gait FSM
• states represent gaits
• each edge has:– geometric preconditions– motion recipe– priority
• self-loops: gait-preserving motion that changes grasps
Complete System
More Results
Future Work
• 3D
• quadrupeds, other characters
• “grasp surfaces”
• non-limb grasping
• add concept of time, speed
• use machine learning
~FIN~
http://www.dgp.toronto.edu/~mac/thesis
Appendix
(extra slides)
Alternate gradient descent view
Smoothing Algorithm
Need for Limb Smoothing
Limb Smoothing Solution
Implemented GFSM
Contributions
• human character animation algorithmfor constrained environments– grasp point discretization of environment– grasp constraint– comfort modeling using heuristics– gait FSM– adapted RPP algorithms to grasp constraint