Reproduce “PADL: Language-Directed Physics-Based Character 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
Jordan Juravsky, Yunrong Guo, Sanja Fidler, Xue Bin Peng
NVIDIA and University of Waterloo researchers developed PADL, a framework enabling natural language control of physically simulated characters. The system learns a joint motion-text embedding space and uses adversarial reinforcement learning to train policies, achieving accurate skill reproduction and dynamic multi-task control in complex environments.
Key insights
- A custom-learned 128D motion-text skill embedding is more effective for conditioning physics-based control policies than raw or PCA-reduced CLIP text embeddings, leading to improved motion quality.
- Dynamic selection of tasks and objects from natural language commands can be achieved by aggregating single-task policies and using a multiple-choice QA model.
- The learned skill embedding space enables semantically meaningful interpolations between different motor skills, producing coherent intermediate behaviors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.