Reproduce “A Self-Correcting Vision-Language-Action Model for Fast and Slow System Manipulation” 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
Chenxuan Li, Jiaming Liu, Guanqun Wang, Xiaoqi Li, Sixiang Chen, Liang Heng, Chuyan Xiong, Jiaxin Ge and 3 more
From Peking University's State Key Laboratory of Multimedia Information Processing, SC-VLA introduces a Vision-Language-Action model that integrates a "fast" system for direct action with a "slow" self-correction system. This framework enables robots to automatically detect and correct low-level SE(3) pose errors, improving manipulation accuracy by 30% on seen and 21% on unseen tasks compared to prior methods in simulation, and achieving a 40% improvement in a real-world stack block task over previous SOTA.
Key insights
- Emulating a human-like "fast and slow" cognitive dual-process system can significantly enhance robotic manipulation robustness and adaptability.
- Direct correction of low-level SE(3) poses, rather than high-level language prompts, is crucial for effective error recovery in robotic manipulation.
- Continuous knowledge transfer from a reflective correction system to a direct action system enables robots to autonomously learn from mistakes and improve their base policy over time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.