robotics
The Robot Monitor Warns Before the Failure
VLA-Tug probes latent layers for calibrated early failure detection and classifies the likely failure mode with little overhead.
Summary
VLA-Tug probes latent layers for calibrated early failure detection and classifies the likely failure mode with little overhead.
Training on the pre-failure segment improves early detection by as much as 30 percent at matched false-positive rates over the strongest learned baseline. Under unseen-task shift, missed-detection and false-alarm bounds remain valid in up to 86 percent of runs, versus at most 21 percent for baselines. Top-three failure-mode accuracy rises by up to 44 percent over a majority baseline, and the authors release 3,180 annotated manipulation episodes. These are benchmarked monitoring results, not a safety certification.
Why it matters
VLA-Tug probes latent layers for calibrated early failure detection and classifies the likely failure mode with little overhead.
Limits and context
- Under unseen-task shift, missed-detection and false-alarm bounds remain valid in up to 86 percent of runs, versus at most 21 percent for baselines.
- These are benchmarked monitoring results, not a safety certification.
Key claims
VLA-Tug probes latent layers for calibrated early failure detection and classifies the likely failure mode with little overhead.
Qualification: Under unseen-task shift, missed-detection and false-alarm bounds remain valid in up to 86 percent of runs, versus at most 21 percent for baselines.
Evidence: source-2026-11-06-008
Sources
- arXiv preprint 2611.08425, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.