Skip to content

← ArchivePaper2024

SuperPADL: Scaling Language-Directed Physics-Based Control with Progressive Supervised Distillation

Jordan Juravsky, Yunrong Guo, Sanja Fidler, Xue Bin Peng

SIGGRAPHAcademic24 citesMotion Synthesis

Progressive distillation from thousands of RL expert policies into a single real-time controller covering over 5,000 language-specified skills.

Abstract

Physically-simulated models for human motion can generate high-quality responsive character animations, often in real-time. Natural language serves as a flexible interface for controlling these models, allowing expert and non-expert users to quickly create and edit their animations. Many recent physics-based animation methods, including those that use text interfaces, train control policies using reinforcement learning (RL). However, scaling these methods beyond several hundred motions has remained challenging. Meanwhile, kinematic animation models are able to successfully learn from thousands of diverse motions by leveraging supervised learning methods. Inspired by these successes, in this work we introduce SuperPADL, a scalable framework for physics-based text-to-motion that leverages both RL and supervised learning to train controllers on thousands of diverse motion clips. SuperPADL is trained in stages using progressive distillation, starting with a large number of specialized experts using RL. These experts are then iteratively distilled into larger, more robust policies using a combination of reinforcement learning and supervised learning. Our final SuperPADL controller is trained on a dataset containing over 5000 skills and runs in real time on a consumer GPU.

How to read this

Category
Method: scalable language-directed physics-based character control
Contributions
  • Introduces SuperPADL, a framework combining RL and supervised learning to scale physics-based text-to-motion to thousands of clips
  • Uses progressive distillation, training many RL experts then iteratively distilling them into larger, more robust policies
  • Yields a single real-time controller covering a large set of language-specified skills
Context
Extends language-directed physics-based control from Juravsky et al.'s PADL, borrowing the scaling intuition of supervised kinematic models to push RL controllers past the usual few-hundred-motion ceiling.Builds on: PADL: Language-Directed Physics-Based Character Control
Correctness
Centers on the assumption that distillation preserves expert skill while improving robustness and coverage; validated on thousands of diverse motion clips, but the abstract is truncated, so read the paper for the exact skill count, metrics, and any motions where distillation degrades fidelity.
Clarity
Reasonably accessible at a high level; a first pass conveys the staged-distillation idea, a second pass is needed for the RL plus supervised loss details.
How to read it
Focus on the staged pipeline (expert RL -> iterative distillation) and why supervised signal is mixed in; do a second pass on the distillation objective and evaluation if you care about reproducing the scaling.

Builds on

Built upon by

Nothing yet.

Related work

Keywords

This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →