Reproduce “Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling” 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
Yuejiang Liu, Jubayer Ibn Hamid, Annie Xie, Yoonho Lee, Maximilian Du, Chelsea Finn
Researchers at Stanford University developed Bidirectional Decoding (BID), a test-time inference algorithm that enhances the performance and robustness of generative behavioral cloning policies by addressing the consistency-reactivity trade-off in action chunking. BID selects optimal action sequences by combining backward coherence with previous decisions and forward contrast against positive and negative policy samples, achieving a 32% relative improvement over baselines in simulation and demonstrating effectiveness in real-world dynamic object manipulation.
Key insights
- Action chunking in robot learning presents an inherent "consistency-reactivity trade-off," where longer action horizons provide consistency but reduce reactivity to environmental changes, and vice-versa.
- Bidirectional Decoding (BID) leverages both "backward coherence" (maintaining consistency with past decisions) and "forward contrast" (selecting optimal future plans by comparing to expert/suboptimal samples) to overcome this trade-off.
- Applying decoding strategies from large language models to low-level robotic control via sample comparison can significantly improve robustness and performance without requiring value functions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.