robotics
The Robot Reviewer Learns When Not to Interrupt
FastHarness schedules model supervision less often as reviewed experience accumulates, then restores scrutiny around failures and recovery.
Summary
FastHarness schedules model supervision less often as reviewed experience accumulates, then restores scrutiny around failures and recovery.
Reviewing every action chunk can make agentic robot supervision expensive even when the policy repeats a familiar stage. FastHarness stores reviewed outcomes and stage experience to lengthen the fast path, while a separate recovery path concentrates oversight after errors. Once recovery is verified, routine execution resumes and the failure remains in memory. Across four simulated manipulation tasks, the authors report a 57.2 percent average decline in reviewer-token use from the first evaluated episode to the mean of the last five, and mean episode time falling from 85.5 to 52.0 minutes. Success rose from 80 percent for the compared GPT-as-policy setup to 97.19 percent. The result uses four simulations, one frozen policy and one reviewer configuration; it does not establish safe sparse review in open-world robotics.
Why it matters
FastHarness schedules model supervision less often as reviewed experience accumulates, then restores scrutiny around failures and recovery.
Limits and context
- The result uses four simulations, one frozen policy and one reviewer configuration; it does not establish safe sparse review in open-world robotics.
Key claims
FastHarness schedules model supervision less often as reviewed experience accumulates, then restores scrutiny around failures and recovery.
Qualification: The result uses four simulations, one frozen policy and one reviewer configuration; it does not establish safe sparse review in open-world robotics.
Evidence: source-2026-10-25-016
Sources
- arXiv preprint 2610.31427, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.