Reproduce “Towards the Harness of Embodied Agents” 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

Qi Wang, Tianyi Wang, Chengyang Li, Shikun Ban, Yurun Chen, Yizhong Ge, Jason Qin, Chengtai Li and 1 more

Thea adapts a coding-agent harness to robot tasks, letting a model select callable capabilities and use feedback to choose actions. A scene graph provides readable world state, while an evaluator checks manipulation outcomes. Thea had the highest success rate among compared systems at three task levels, but evaluator errors included false successes, so the results do not establish perfect verification.

Key insights

  • Reliability in embodied AI can be substantially enhanced by a system-level "harness" that orchestrates capabilities and enables closed-loop evaluation and recovery.
  • Explicitly addressing the "readability of the world" via structured representations (scene graph) and "verifiability of outcomes" via independent evaluation is crucial for physical agents.
  • The modular design allows for additive growth of robot capabilities and portability across different robot embodiments through interchangeable profiles and tools.

Summary by alphaXiv.

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