DynaHarness: A Dynamic Physical Harness for Self-Evolving Robot Agents
DynaHarness couples semantic reasoning with physical governance in robot agents, raising LIBERO-Pro success from 17.5% to 75.2% over a frozen VLA policy.
DynaHarness is a dynamic physical harness that pairs a slow semantic reasoning brain with a fast grounding-and-monitoring brain through a shared execution contract, turning failure evidence into validated capability revisions. Failure attribution localizes faults across analytic skills, recovery skills, and the frozen VLA, with paired regression checks governing admission of revisions. On LIBERO-Pro it achieves 75.2% success on 800 newly sampled initial states versus 17.5% for the frozen policy, and full dynamic execution reaches 74.0% versus 63.9% under nominal one-step replanning.
- Slow brain proposes capabilities; fast brain grounds, monitors, and refuses unresolved actions.
- Failure attribution localizes faults and directs targeted capability revisions.
- 75.2% success on LIBERO-Pro across 800 new initial states versus 17.5% frozen policy.
- Full dynamic execution reaches 74.0% versus 63.9% with one-step replanning.
From the original article
1 min read at arxiv.orgPretrained robot policies provide useful action priors, but long-horizon manipulation still requires coordination between semantic reasoning and physical execution. Semantic reasoning operates at a coarser timescale than physical interaction, while episode-level failures provide limited guidance on which system component should be revised.
By Haoyuan Deng · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.