Reproduce “A Few Words Go a Long Way: Language Guided Robot Policy Synthesis” 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

Daphne Chen, Archit Ritesh Jain, Eric Goossen, Emma Romig, Michael Murray, Nick Walker, Maya Cakmak

ARCHITECT, a framework for robot policy acquisition, integrates interactive program synthesis guided by natural language and human feedback with a persistent skill library. This approach enables more adaptable and data-efficient robot control, outperforming state-of-the-art VLA baselines and significantly reducing human intervention in new tasks by amortizing human effort.

Key insights

  • Interactive program synthesis, when guided by iterative human language corrections, can effectively bridge the 'specification gap' in modular robotics, allowing for granular policy refinement.
  • A persistent skill library enables long-term learning and knowledge transfer, significantly reducing human intervention for new, related tasks by accumulating reusable behaviors.
  • Human corrections are superior to VLM-generated feedback for diagnosing subtle physical and perceptual failures that current automated vision models struggle to detect or infer.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.