DynaHarness: A Dynamic Physical Harness for Self-Evolving Robot Agents

2026-09-30

AI Summary

This study addresses the challenges of coordinating semantic reasoning with physical execution and identifying failure causes in long-horizon robotic manipulation. We propose a dual-brain dynamic physics framework that couples slow semantic planning via Vision-Language-Action (VLA) models with fast physical monitoring, achieving closed-loop control through a shared execution contract mechanism. Furthermore, we introduce a paired regression checking algorithm to automatically verify and revise capability libraries, thereby enabling autonomous system self-evolution. Evaluated on the LIBERO-Pro benchmark, our method achieves a success rate of 75.2%, substantially outperforming the frozen-policy baseline at 17.5%. These results demonstrate a dual breakthrough in both robustness and adaptive autonomy for long-horizon manipulation tasks.

Abstract

Pretrained 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. We propose DynaHarness, a dynamic physical harness that couples semantic reasoning with physical governance through a shared execution contract and turns failure evidence into validated capability revisions. To be more specific, the slow brain proposes capabilities and symbolic arguments, while the fast brain grounds and monitors commands, refuses unresolved actions, substitutes capabilities, and requests replans when needed. The physical execution contract bounds each accepted command and records execution evidence across analytic skills, recovery skills, and the frozen VLA. Failure attribution localizes faults in these records and directs targeted revisions of reusable capabilities or execution mechanisms. Paired regression checks govern admission or rejection, closing the self-evolution loop. On LIBERO-Pro, DynaHarness achieves 75.2% on 800 newly sampled initial states, compared with 17.5% for the frozen policy. With the same capability library, full dynamic execution reaches 74.0% versus 63.9% under nominal one-step replanning. This demonstrates the value of DynaHarness as a dynamic physical harness that governs how existing capabilities are grounded, monitored, and coordinated during execution. Our project page is at https://denghaoyuan123.github.io/Dynaharness_page/.

Problem

• long-horizon manipulation
• semantic reasoning
• physical execution
• failure attribution
• self-evolving robot agents

Innovation

• Self-Evolving Robot Agents
• Dynamic Physical Harness
• Semantic-Physical Coordination
• Failure Attribution
• Vision-Language-Action (VLA)