Reproduce “Self-Evolving Embodied Agents via Skill-Harness Evolution” 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
Peidong Wang, Zhiming Ma, Ying Chang, Xufang Luo, Yiqun Zhang, Zihan Wang, Xiaocui Yang, Shi Feng and 2 more
SHAPER adapts a frozen embodied agent by evolving a textual skill and a context-code harness that selects and formats execution history. Rollouts are judged, summarized, and used by an optimizer while the planner and executor are fixed. On a 231-question ESI-Bench subset, SHAPER reaches 49.8% micro accuracy versus 32.5% for Seed Agent, without evidence of real-robot or cross-embodiment validation.
Key insights
- Procedural guidance and context construction address different failure modes: skill evolution improves command selection, decomposition, progress checks, and recovery, while harness evolution improves which visual and action-history evidence reaches the planner.
- Joint evolution is more effective than either component alone, particularly under task and category shifts; on VLABench, full SHAPER reaches 34.50% versus 33.50% for skill-only and 30.50% for harness-only evolution.
- Context management can be as important as planning: retaining diverse visual keyframes and deterministic crops recovered information needed for mirror-reflection reasoning on ESI-Bench.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.