Download presentation
Presentation is loading. Please wait.
1
Simulating Biped Behaviors from Human Motion Data
Kwang Won Sok Manmyung Kim Jehee. Lee
2
Motion graph v.s. robotics controller
Goal Dynamics simulation Learn various motor skills from human Controllable Motion graph v.s. robotics controller
3
Challenge Controller Hard to design Unnatural Mocap Imprecise data
Highly simplified model
4
Approach Rectify mocap data to be physically feasible
Good data v.s. good method “Controller” -- “state-action” pair A set of desired trajectories
5
Interpolation and refinement
Controller Learning Data Rectification Interpolation and refinement
6
Data Rectification Space-time Optimization
Variables: displacement over input data Objective: simulation result be close to input data [Constraints: forward simulation with simple PD control]
9
Interpolation and Refinement
Feature (state) vector Joint angles and velocities Root position, velocity and direction (spine) Foot position and velocity Ground contact ( ON/OFF ) Blend to a successful motion when failure happens Enrich the existing database Dimensionality Is success guaranteed?
10
Transitions Rectify motion graph built on the enriched database
11
Result Balance
12
Result Transition between various motions
13
Questions Your impression Interesting or dumb? What’s new
motion graph, reinforcement learning? Why not 3D Curse of dimensionality? What makes it into SIGGRAPH
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.