SAFE: Multitask Failure Detection for Vision-Language-Action Models

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

SAFE is a multitask failure detector for generalist Vision-Language-Action (VLA) models that probes internal latent features to predict a scalar failure score, generalizing zero-shot to unseen tasks. It exploits the empirical finding that failed rollouts cluster in a task-agnostic "failure zone" in the VLA feature space. Using functional conformal prediction for threshold calibration, SAFE achieves state-of-the-art failure detection accuracy and latency trade-off across OpenVLA, π0, and π0-FAST on both simulation and real robots.

Problem

Existing failure detectors are trained and evaluated per-task, making them unsuitable for generalist VLAs that routinely encounter novel task instructions at deployment. VLAs achieve only 30–60% success on unseen tasks out-of-the-box, yet no prior work demonstrated zero-shot multitask failure detection—i.e., a single detector trained on seen tasks that generalizes to held-out tasks without any additional rollouts or fine-tuning.

Method

SAFE operates in three stages:

1. Feature extraction: At each inference timestep \(t\), extract the hidden state vector \(e_t\) from the last layer of the VLA (before token logit decoding or diffusion velocity field output).

2. Failure score prediction: Two backbone options:

3. SAFE-MLP: Independent per-timestep MLP \(g(e_t)\) with sigmoid activation; failure score accumulates as \(s_t = \sum_{\tau \leq t} \sigma(g(e_\tau))\); trained with L1 loss pushing scores up for failures, down for successes.

4. SAFE-LSTM: Sequential LSTM over \(e_{0:t}\); outputs \(s_t = \sigma(\text{LSTM}(e_{0:t}))\); trained with binary cross-entropy. Both backbones are kept shallow (1–2 layers) to avoid overfitting.

5. Threshold calibration via functional conformal prediction (CP): A time-varying upper band \(\delta_t = \mu_t + h_t\) is calibrated on successful rollouts from seen tasks. If \(s_t > \delta_t\) at any timestep, a failure flag is raised with nominal false-positive rate \(\leq \alpha\).

The motivation is a t-SNE visualization showing that VLA internal features from different tasks converge to the same "failure zone" when the policy is failing, enabling cross-task generalization.

Key Contributions

• Formal introduction of the multitask failure detection problem for generalist VLAs: train on seen tasks, evaluate zero-shot on unseen tasks.
• Empirical analysis demonstrating task-agnostic failure clustering in VLA latent space (shown for π0-FAST on LIBERO-10).
• SAFE: a lightweight (2.3M parameter LSTM; <1ms overhead, <1% of VLA inference time) failure detector trained on multi-task rollouts that outperforms all baselines on unseen tasks.
• Comprehensive evaluation across three VLA architectures (OpenVLA, π0, π0-FAST), two simulation benchmarks (LIBERO-10, SimplerEnv), and two real-robot platforms (Franka+ZED, WidowX).
• Adaptation of LLM uncertainty quantification methods (token probability, semantic entropy) as VLA failure detection baselines.

Results

• Simulation ROC-AUC (unseen tasks, averaged across benchmarks): SAFE-MLP and SAFE-LSTM outperform the best non-sampling baseline by 4–5% on unseen tasks.
• OpenVLA LIBERO unseen: SAFE-LSTM 92.98, SAFE-MLP 90.06 vs. best baseline Cosine k-NN 92.09.
• π0-FAST LIBERO unseen: SAFE-LSTM 88.85, SAFE-MLP 89.50 vs. best baseline Avg entropy 90.19.
• π0 LIBERO unseen: SAFE-LSTM 77.04, SAFE-MLP 78.00 vs. best baseline Cosine k-NN 73.93.
• π0* SimplerEnv unseen: SAFE-LSTM 84.48, SAFE-MLP 80.44 vs. best baseline STAC 85.31.
• Real-world ROC-AUC (unseen tasks):
• π0-FAST Franka: SAFE-MLP 64.16 (best); SAFE-LSTM 58.70 vs. Euclid. k-NN 60.27.
• OpenVLA WidowX: SAFE-MLP 88.42 (best); SAFE-LSTM 71.80 vs. Mahala. dist. 70.00.
• Efficiency: SAFE adds <1ms latency vs. 149ms for π0 inference; STAC with 10 samples is 152–221% slower than single-sample inference.
• Detection time: Functional CP curves show SAFE detects failures earlier (lower T-det) at equivalent balanced accuracy compared to most baselines on OpenVLA-LIBERO and π0-SimplerEnv.

Limitations

• Evaluated only on manipulation tasks; generalization across embodiments, sim-to-real transfer, or action-less video VLAs is untested.
• SAFE uses only the final layer features; multi-layer aggregation strategies remain unexplored.
• Requires collecting successful and failed rollouts on seen tasks before deployment; not truly zero-data.
• Sampling-based baselines (STAC with 256 samples) occasionally outperform SAFE on specific benchmarks (e.g., STAC leads on π0-FAST LIBERO unseen) but are excluded from real-robot comparisons due to inference cost.
• Conformal prediction calibration relies on the exchangeability assumption, which may not hold under significant distribution shift between seen and unseen tasks.

Relevance to Vision-Language Models

SAFE directly addresses the reliability gap of deployed VLMs-turned-VLAs—large multimodal models repurposed for robot control—by showing their internal representations already encode task-agnostic success/failure signals transferable across novel instructions, a property inherited from their VLM pretraining. This connects to the broader literature on probing LLM/VLM latent spaces for hallucination and uncertainty detection, extending those ideas from text generation to continuous action generation. For VLM researchers, the key finding is that the semantic richness of VLA features (learned from internet-scale data) enables zero-shot behavioral monitoring without task-specific supervision. This has direct implications for safe deployment of any instruction-following multimodal model in open-world interactive settings.