SuperTrack: Motion Tracking for Physically Simulated Characters Using Supervisory Signals

How to read this

Category

Method: learning-based motion tracking for physics-based characters

Contributions

• Frames physics-based motion tracking as supervised learning, optimizing the control policy directly via back-propagation
• Trains a world model that approximates the environment's transition function, acting as a differentiable physics simulator
• Reports higher control quality in shorter training time with reduced sensitivity to dataset size and experience-gathering rate versus PPO-based methods

Context

Positioned against model-free PPO controllers for simulated characters and builds on data-driven responsive control (Bergamin et al.'s DReCon).Builds on: DReCon: Data-Driven Responsive Control of Physics-Based Characters

Correctness

Key assumption is that a learned world model is an accurate-enough differentiable proxy of physics for policy gradients; comparisons are made primarily to PPO on motion tracking, so the world-model approximation error and behavior outside the trained motion distribution are the limitations to watch.

Clarity

The supervised-versus-model-free framing is clear; the world-model training and differentiable-optimization mechanics need a second pass.

How to read it

First pass for the world-model-as-differentiable-simulator idea and why it beats PPO here; second pass on the training loop and loss if you want to reproduce the control quality.