robotics
Thirty-Four Thousand Parameters Read Failure From a Frozen Robot Model
A lightweight readout reached 85.68 AUROC across seven source tasks and added 0.2256 milliseconds after the predictive state existed.
Summary
A lightweight readout reached 85.68 AUROC across seven source tasks and added 0.2256 milliseconds after the predictive state existed.
FARM asks whether a frozen robotic world model already carries usable signs of impending failure. A supervised readout with 33,985 parameters maps its predictive states to stepwise failure scores and trajectory risk without updating the backbone. Five-fold evaluation across seven tasks reached pooled AUROC of 85.68 and AUPRC of 88.59, then fixed and readout-only adaptation were tested across PIPER X, SO-101 and Franka robot populations. Once the frozen state was available, mean added CUDA latency was 0.2256 milliseconds. These are predictive monitoring results, not proof that the system safely stops or recovers a robot.
Why it matters
A lightweight readout reached 85.68 AUROC across seven source tasks and added 0.2256 milliseconds after the predictive state existed.
Limits and context
- Five-fold evaluation across seven tasks reached pooled AUROC of 85.68 and AUPRC of 88.59, then fixed and readout-only adaptation were tested across PIPER X, SO-101 and Franka robot populations.
- These are predictive monitoring results, not proof that the system safely stops or recovers a robot.
Key claims
A lightweight readout reached 85.68 AUROC across seven source tasks and added 0.2256 milliseconds after the predictive state existed.
Qualification: Five-fold evaluation across seven tasks reached pooled AUROC of 85.68 and AUPRC of 88.59, then fixed and readout-only adaptation were tested across PIPER X, SO-101 and Franka robot populations.
Evidence: source-2026-09-11-013
Sources
- arXiv preprint 2609.11445arXiv · primary research
Corrections
No corrections have been recorded for this story.