Reproduce “Temporal Action Selection for Action Chunking” 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
Yueyang Weng, Xiaopeng Zhang, Yongjin Mu, Yingcong Zhu, Yanjie Li
Researchers from Harbin Institute of Technology Shenzhen introduce Temporal Action Selection (TAS), an algorithm that dynamically selects actions from a cache of temporally diverse predictions, to balance reactivity and decision consistency in robotic manipulation. TAS improves average success rates by 35-42% across various tasks and policy architectures, and significantly enhances the training efficiency and performance of residual reinforcement learning in real-world deployments.
Key insights
- Caching action predictions from distinct past timesteps is crucial for enabling the selector to balance immediate reactivity and long-term decision consistency, outperforming synchronous candidate generation.
- The Implicit Space-Aware Selector's structured inductive bias, using vector similarity for scoring, provides superior generalization and avoids overfitting compared to MLP-based selectors.
- The selector dynamically adapts its preference, choosing more reactive actions during precision-critical task phases (e.g., grasping) and more consistent actions during less demanding phases (e.g., transport).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.