DynaHarness: A Dynamic Physical Harness for Self-Evolving Robot Agents
DynaHarness reports a 75.2% LIBERO-Pro success rate by letting a robot policy revise capabilities from execution failures.
The paper frames the system as a “slow brain” for semantic capability proposals and a “fast brain” for grounding, monitoring, refusals, substitutions, and replans. Its execution contract records evidence from analytic skills, recovery skills, and a frozen VLA, then uses failure attribution to target revisions. Paired regression checks decide whether those revisions are admitted. The authors compare 75.2% on 800 newly sampled initial states with 17.5% for the frozen policy, and 74.0% full dynamic execution against 63.9% nominal one-step replanning. ArXiv · AI/CL/LG's note
The paper frames the system as a “slow brain” for semantic capability proposals and a “fast brain” for grounding, monitoring, refusals, substitutions, and replans. Its execution contract records evidence from analytic skills, recovery skills, and a frozen VLA, then uses failure attribution to target revisions. Paired regression checks decide whether those revisions are admitted. The authors compare 75.2% on 800 newly sampled initial states with 17.5% for the frozen policy, and 74.0% full dynamic execution against 63.9% nominal one-step replanning. ArXiv · AI/CL/LG's note
score 4