Streaming Failure Forecasting: VLA Failure Monitors That Keep Learning After Deployment
Streaming Failure Forecasting: VLA Failure Monitors That Keep Learning After Deployment
This research presents Streaming Failure Forecasting (SFF), a method for vision-language-action (VLA) policies that maintains failure monitoring accuracy after deployment by refitting only the final linear layer of its network using success/failure labels from each episode. Unlike traditional approaches that freeze monitors offline, SFF adapts to unannounced task switches and policy updates on real SO-101 arms and in simulation, ranking failures better than frozen or reset-based baselines while operating at 2.18 ms per call on a Jetson CPU core.
The paper introduces Streaming Failure Forecasting (SFF), a method that refits only the final linear layer of a failure monitor after each episode to adapt to task or policy changes in VLA systems, achieving better failure ranking than frozen monitors on simulated drift and real-robot runs.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
Source：arXiv Robotics — research abstracts · arxiv.org