Reproduce “ADAPT: Agile Diffusion Action Priors for Robust and Steerable Online Text-Driven Humanoid Control” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yan Wu, Chenhao Li, Kaifeng Zhao, Gen Li, Marco Hutter, Siyu Tang

ADAPT controls a humanoid from recent physical state and changing text commands, then adds residual correction for execution robustness. Frame-level skill labels train the controller, while a separate noise-steering policy adapts its frozen diffusion prior for goal reaching. In prompt-switching tests, no-fall success rises from 0.804 to 0.984, while R@1 falls; goal-reaching success reaches 97.1% in the tested setup.

Key insights

  • Combining a diffusion-based behavior prior with a spatially constrained residual RL policy effectively balances robustness and semantic preservation, enabling stable real-time interactive control for humanoids.
  • Fine-grained, skill-level language supervision, derived from datasets like BABEL, is critical for training policies capable of generating responsive and smooth transitions between diverse whole-body humanoid behaviors.
  • A frozen diffusion action prior can be robustly adapted for downstream tasks by steering its latent noise input with a lightweight task-conditioned RL policy, preserving language-defined motion styles.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.