Reproduce “Strategy and Skill Learning for Physics-based Table Tennis Animation” 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
Jiashun Wang, Jessica Hodgins, Jungdam Won
This research introduces a hierarchical control system for physics-based table tennis animation, enabling virtual characters to learn diverse motor skills and dynamic strategic decision-making. The system effectively counters mode collapse in skill learning and facilitates realistic real-time human-agent interaction in virtual reality, demonstrated by improved rally sustainability and competitive win rates against human and AI opponents.
Key insights
- A hierarchical control architecture effectively decouples strategic planning from low-level motor execution, allowing for specialized learning in each domain.
- The mixer policy, by blending actions from a universal policy and skill-specific policies, is crucial for mitigating mode collapse and ensuring fluid transitions between diverse motor skills.
- Using iterative behavior cloning with a Conditional Variational Autoencoder (CVAE) allows the strategy controller to learn robust and diverse decision-making policies from various interaction data, including human play.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.