Reproduce “Beyond Success: Refining Elegant Robot Manipulation from Mixed-Quality Data via Just-in-Time Intervention” 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
Yanbo Mao, Jianlong Fu, Ruoxuan Zhang, Hongxia Xie, Meibao Yao
A framework refines robot manipulation elegance from mixed-quality demonstration data by introducing a Just-in-Time Intervention (JITI) mechanism that guides pre-trained Vision-Language-Action (VLA) policies. The method boosts Elegant Success Rate (ESR) by 17.4 to 21.2 percentage points in simulation and 23.7 percentage points in real-world tasks, improving execution quality without altering the base VLA model.
Key insights
- Formalizing and explicitly rewarding 'elegant execution' via Implicit Task Constraints (ITCs) allows for a more nuanced evaluation and improvement of robotic manipulation quality beyond mere task success.
- A non-invasive, lightweight Elegance Critic, trained with dense, task-specific elegance rewards, can effectively guide and refine large, pre-trained VLA policies without modifying their core parameters.
- Just-in-Time Intervention (JITI) offers an efficient balance between performance gains and computational cost by providing targeted guidance only at critical junctures, enabling selective high-quality action selection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.