Failure Prediction and Classification in Vision-Language-Action Models Via Calibrated Latent Probing
arXiv cs.AI · 6 Nov 2026, 04:21 AEST
Vision-Language-Action models (VLAs) face challenges in performance under distribution shifts, necessitating effective runtime monitoring. The introduction of VLA-Tug enhances early failure detection and reduces false-alarm rates with minimal inference overhead.
Key facts
- Deploying VLAs in real-world systems therefore requires a runtime monitor that is timely, calibrated, and informative while being efficient enough to run on the edge.
- We present VLA-Tug, a lightweight probe over the latent layers of a VLA backbone that improves early failure detection and bounds the missed-detection and false-alarm rates, with minimal inference overhead.
- We show that supervising on the pre-failure segment of a trajectory increases early failure detection at matched false-positive rates by up to 30% over the strongest learned baseline.
- Under shift to unseen tasks, the bounds on missed-detection and false-alarm rates remain valid in up to 86% of runs, versus at most 21% for baselines.
- VLA-Tug predicts the failure mode to enable corrective actions, raising top-3 accuracy on unseen tasks by up to 44% over a majority baseline.
Read aloud
Vision-Language-Action models VLAs face challenges in performance under distribution shifts, necessitating effective runtime monitoring. The introduction of VLA-Tug enhances early failure detection and reduces false-alarm rates with minimal inference overhead. Deploying VLAs in real-world systems therefore requires a runtime monitor that is timely, calibrated, and informative while being efficient enough to run on the edge. We present VLA-Tug, a lightweight probe over the latent layers of a VLA backbone that improves early failure detection and bounds the missed-detection and false-alarm rates, with minimal inference overhead. We show that supervising on the pre-failure segment of a trajectory increases early failure detection at matched false-positive rates by up to thirty percent over the strongest learned baseline. Under shift to unseen tasks, the bounds on missed-detection and false-alarm rates remain valid in up to eighty six percent of runs, versus at most twenty one percent for baselines. VLA-Tug predicts the failure mode to enable corrective actions, raising top-three accuracy on unseen tasks by up to forty four percent over a majority baseline.
Sources in this cluster
- Failure Prediction and Classification in Vision-Language-Action Models Via Calibrated Latent Probing
arXiv cs.AI · 6 Nov 2026, 04:21 AEST