Research map / Vision-language and multimodal / Robot action and navigation
SAFE: Multitask Failure Detection for Vision-Language-Action Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SAFE, a multitask failure detector that probes VLA internal features with a small MLP or LSTM trained on successful and failed rollouts, and calibrates alerts with functional conformal prediction, evaluated on unseen tasks across OpenVLA, pi0, and pi0-FAST.
Problem
Detecting failures of generalist vision-language-action policies on novel, unseen tasks without collecting rollouts or finetuning the detector per task, in a timely and efficient manner.
Approach
VLA internal features separate successful and failed rollouts into a task-generic "failure zone"; a lightweight MLP or LSTM probe on these features outputs a per-timestep failure score, and functional conformal prediction sets a time-varying threshold calibrated on seen tasks.
Key results
- SAFE-MLP averages 78.00% ROC-AUC on unseen tasks across simulation benchmarks, outperforming the best baseline by about 4-5 points.
- SAFE-LSTM and SAFE-MLP beat baselines on OpenVLA+LIBERO and pi0*+SimplerEnv in the accuracy vs. detection-time trade-off.
- SAFE adds under 1 ms overhead (SAFE-LSTM 2.3M params, 0.73 ms) versus pi0's 149 ms inference, while multi-sample baselines are much slower.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAFE-MLP instead of the strongest average baseline | Simulation failure detection · Simulation benchmarks, average Unseen | ROC-AUC: 73.93 -> 78.00 | Helped |
| Use SAFE-LSTM instead of the strongest average baseline | Simulation failure detection · Simulation benchmarks, average Seen | ROC-AUC: 81.28 -> 82.26 | Helped |
| Use SAFE-MLP instead of action total variance | OpenVLA failure detection · LIBERO Unseen | ROC-AUC: 65.43 -> 73.47 | Helped |
| Use SAFE-MLP instead of action total variance | OpenVLA failure detection · LIBERO Seen | ROC-AUC: 62.76 -> 72.68 | Helped |
| Use SAFE-LSTM instead of Mahalanobis distance | pi_0-FAST failure detection · LIBERO Seen | ROC-AUC: 93.56 -> 92.98 | Hurt |
| Use SAFE-LSTM instead of cosine distance k-NN | pi_0-FAST failure detection · LIBERO Unseen | ROC-AUC: 84.64 -> 84.48 | Hurt |
| Use SAFE-MLP instead of gripper total variance | pi_0 failure detection · LIBERO Unseen | ROC-AUC: 75.19 -> 73.27 | Hurt |
| Use SAFE-LSTM instead of action total variance | pi_0 failure detection · LIBERO Seen | ROC-AUC: 77.20 -> 76.98 | No clear effect |
| Use SAFE-MLP instead of cosine distance k-NN | pi_0* failure detection · SimplerEnv Seen | ROC-AUC: 90.19 -> 89.50 | Hurt |
| Use SAFE-MLP instead of cosine distance k-NN | pi_0* failure detection · SimplerEnv Unseen | ROC-AUC: 71.32 -> 84.82 | Helped |
| Use SAFE-MLP rather than SAFE-LSTM (ablation) | Simulation failure detection · Simulation benchmarks, average Unseen | ROC-AUC: 77.04 -> 78.00 | Helped |
| Use SAFE-LSTM rather than SAFE-MLP (ablation) | Simulation failure detection · Simulation benchmarks, average Seen | ROC-AUC: 81.43 -> 82.26 | Helped |
Surprising findings
- Real-world Franka tasks with pi0-FAST yield very low detection performance (ROC-AUC at most 64 on unseen tasks), with failure embeddings not clearly separable.
- Token uncertainty baselines perform poorly, close to chance in many settings.
Limitations
- Only manipulation tasks considered; generalization across embodiments, sim-to-real, or action-less videos untested.
- Uses only last-layer VLA features; multi-layer aggregation left open.
- Conformal prediction guarantees weaker in multitask setting since calibration and evaluation distributions differ, so TNR sometimes falls below 1-alpha.
Experimental setup
Datasets: LIBERO-10 SimplerEnv DROID
Benchmarks: LIBERO-10 SimplerEnv Real Franka
Metrics: ROC-AUC balanced accuracy TPR FPR detection time
Baselines: Max prob. token uncertainty Mahalanobis distance Cosine distance k-NN PCA-KMeans RND LogpZO STAC STAC-Single Action total var. Cluster entropy
Builds on: functional conformal prediction LogpZO STAC LSTM
Closest papers by meaning
- AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation (ICLR 2025 · similarity 0.92)
- Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning (NeurIPS 2025 · similarity 0.91)
- Self-Refining Vision Language Model for Robotic Failure Detection and Reasoning (ICLR 2026 · similarity 0.90)
- Vision-Language Models for Robot Success Detection (AAAI 2024 · similarity 0.88)
- CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.