Near-Optimal Character Animation with Continuous Control

How to read this

Category

Method: data-driven near-optimal character controllers

Contributions

• Automatic computation of near-optimal real-time controllers from a motion-capture corpus and a desired task.
• A motion engine that blends precaptured clips and prevents foot-skate without inverse kinematics.
• A control policy using a compact value function as a linear combination of basis functions, learned via a linear-programming approximation to dynamic programming, plus switchability and separability to manage dimensionality.

Context

Builds on motion-graph style clip-sequencing (Kovar et al.'s Motion Graphs) and frames clip selection as approximate dynamic programming over a low-dimensional value-function basis.Builds on: Motion Graphs

Correctness

Demonstrated on navigation, spinning navigation, and fixed and moving obstacle avoidance (extending to simple crowds); the compactness argument rests on the value functions being smooth for these tasks, so tasks with non-smooth objectives may need more basis functions.

Clarity

Accessible in motivation but reinforcement-learning-flavored; a first pass conveys the basis-function value-function idea, a second pass for the LP approximation and the switchability/separability constructs.

How to read it

First pass for the problem framing and the motion-engine/policy split; second pass on the value-function approximation and dimensionality tricks if you work on learned controllers.