Failure Detection for VLA Models

Vision-Language-Action (VLA) models can perform diverse manipulation tasks directly from visual observations and natural language instructions, but they remain prone to mistakes and unpredictable failures. Deploying these policies without continuous human supervision will require robust runtime monitors that detect when the policy has gone off-course. Existing monitors use external foundation models, such as vision-language models (VLMs) to supervise the policy, introducing significant complexity and computational cost for monitoring at runtime. In this work, we investigate an alternate approach: can a VLA's own internal activations provide the information needed to monitor its execution? Our method trains linear probes on policy activations from demonstrations to estimate progress, then flags executions that deviate from expected progress. In experiments with the open-source VLA $π_{0.5}$ and across simulated and real-world manipulation tasks, this approach is competitive with or outperforms VLM-based monitors in task failure detection, as well as methods which require failure data for training. These findings suggest that simple probes of VLA internals can be used to develop reliable runtime monitors, opening a path toward building policies that can act reliably and monitor their execution without external foundation models.

Benchmarks expose vision-language-action (VLA) policies to few canonical instructions, while exhaustive deployment testing is impossible. We introduce Object-Grounded Attention Monitoring (OGAM), connecting systematic testing to runtime assurance: testing reveals attention divergence between successful and failed executions, and OGAM uses this signal to stop failures beyond the finite suite. We generate scene-grounded instructions through pairwise combinations of action templates and objects, and separately test meaning-preserving paraphrases. All 87 out-of-benchmark cases reveal problematic behavior across OpenVLA, OpenVLA-OFT, UniVLA, and $π_{0.5}$: none completes any of the 24 feasible instructions, while infeasible or hazardous requests also trigger behavior substitution. At each action query, we project gradient-weighted visual attention through object masks and group it by instruction role for comparison across tasks and policies. Dynamic time warping aligns this course with a successful reference despite speed differences; conformal calibration on successful episodes sets the early-stopping threshold for sustained deviations, with a nominal false-stop target of $α=0.05$. Across four policies, OGAM stops 87-100% of failed episodes at median times of 5-12s within a 20s budget, with observed false-stop rates of 3-5%, without failure-labeled training. Finite testing thus identifies attention patterns that support online intervention before failure fully unfolds.

Vision-language-action (VLA) models map visual observations and natural-language instructions to robotic actions, but distribution shifts can compromise their reliability. Because these models may still succeed under out-of-distribution (OOD) conditions, detecting OOD inputs alone is insufficient to predict execution failure. In this paper, we introduce VLA-Scope, a two-stage framework that combines input-shift characterization with execution history to predict failure during OOD rollouts. The first stage uses pooled image and language representations to detect OOD inputs and classify their shift categories. For inputs flagged as OOD, the second stage combines the predicted category, action-prefix features, and execution progress features. A logistic regression model shared across shift categories updates failure risk as execution proceeds. We evaluate the framework with OpenVLA on ten LIBERO-Spatial tasks using leave-one-group-out cross-validation. OOD detection achieves a ROC-AUC of 0.9454, and shift classification achieves 91% accuracy. Evaluated independently of the OOD gate on all 1,400 OOD rollouts, the failure predictor achieves a ROC-AUC of 0.8497 after 60 executed actions, compared with 0.7906 without execution progress features. It also achieves a higher ROC-AUC than the evaluated ActProbe and SAFE-MLP baselines. These results suggest that combining action features with temporally aggregated execution step representations improves failure prediction under input shifts.

Vision-language-action (VLA) policies have shown strong potential for general-purpose robotic manipulation, but they can still fail unpredictably during long-horizon execution, making reliable failure detection essential for safe deployment. Existing methods either rely on visual models that typically detect failures only after erroneous actions have occurred, or use lightweight proactive detectors trained on VLA internal representations. However, these proactive methods are often supervised with trajectory-level labels, causing normal pre-failure behavior in unsuccessful trajectories to be incorrectly labeled as failure. This supervision mismatch introduces label noise and limits both trajectory-level detection accuracy and precise timestamp-level failure localization. In this work, we study fine-grained timestamp-level VLA failure detection while addressing the cost of dense annotation. We propose a data-efficient framework that first leverages unlabeled VLA action chunks to construct action-derived weak supervision signals, capturing abnormal patterns such as inconsistent consecutive chunks, frozen or idle actions, and aggressive random motions. We then use active learning to select only the most uncertain trajectories for timestamp-level annotation and fine-tune the detector with these informative labels. Experiments across multiple VLA policies show that our method improves both timestamp-level and trajectory-level failure detection performance.

Diffusion-based vision-language-action (VLA) policies can generate plausible actions even when their predictions are weakly grounded in the visual and language evidence defining the task. We introduce GUARD, a test-time failure detection method that measures this grounding without modifying the pretrained policy. GUARD estimates the influence of token-indexed entries in the final vision-language model key-value (KV) cache, constructs counterfactual caches by ablating salient KV entries, and compares their denoising responses with the original conditioning. Based on the comparison, we derive GUARD diagnostic stream including sensitivity, attention entropy, modality bias, and grounding efficiency, which are calibrated online and processed by a lightweight temporal classifier. We evaluate GUARD under task-held-out splits across five policy-benchmark settings, using Pi0, SmolVLA, and Alpamayo-1.5 on LIBERO, SimplerEnv, MetaWorld, and PhysicalAI-AV. GUARD achieves the best ROC-AUC on four of five unseen-task settings and ranks second on the remaining setting, improving the average unseen-task ROC-AUC by 5.73 percentage points over the strongest competing runtime monitor while remaining within 0.19 points of the best seen-task average. These results show that directly probing action-head dependence on multimodal evidence provides a transferable failure signal across policies, tasks, embodiments, and domains.

Vision-language-action policies often fail under deployment-time distribution shifts such as clutter, distractor objects, lighting changes, novel objects, altered initial states, and reworded instructions. Hidden-state-based risk probes combined with functional conformal prediction can detect rollout failures, but their reliability depends on calibration data matching deployment conditions. We introduce SAFECAST, which leverages contrast set perturbations to improve hidden-state probe training and calibration for deployment time shift. SAFECAST statistically significantly improves failure detection ROC-AUC scores over a state of the art baseline in both real-world DROID and LIBERO simulation experiments across multiple VLM backbones. We further find that SAFECAST benefits most when both visual and language contrast set perturbations are used to augment data, and that with contrast set perturbations, sim-to-real calibration leads to better probes than using real rollout data only.

Flow matching (FM) has become a popular action head paradigm for modern embodied models. However, as a conditional generative model, it does not explicitly expose its inherent uncertainty, producing faulty action chunks even when it misinterprets the scene or encounters out-of-distribution (OOD) inputs. Therefore, determining when an FM-generated action can be trusted is essential for safe deployment, yet existing uncertainty estimation methods on real-time control suffer from several issues: extra training budget, high computational overhead, and low generalization ability. In this work, we provide a geometric interpretation of FM uncertainty in the velocity field, showing that uncertainty manifests as deviation from an ideal affine-isotropic contraction field. Building on this observation, we introduce denoising acceleration ($\mathrm{accel}$), a highly-generalizable and cost-free uncertainty proxy that measures the bending of the denoising trajectory from a single forward pass, without additional model evaluations, training, or resampling. We theoretically and empirically demonstrate that $\mathrm{accel}$ is a faithful proxy for FM uncertainty and further test its utility in online failure detection. Results show that $\mathrm{accel}$ identifies failing rollouts well before termination, matching or even outperforming costly resampling- and training-based baselines across settings under realistic deployment budget. Code and demos available at: https://github.com/rrrrrrzy/fm-geometry.

Vision Language Action models combine perception, language grounding, and control in a single policy, but their failures are hard to diagnose once visual conditions shift. We test whether OpenVLA feedforward activations contain linearly decodable information about near term task failure in LIBERO manipulation rollouts. The policy is fixed throughout. We log internal activations during execution and fit lightweight monitors after the rollouts are collected. Occlusion is the main controlled stress test. It reduces OpenVLA success from $57\%$ to $17\%$ over $100$ episodes per condition. Under this shift, a logistic probe at layer 16 reaches AUROC $0.972$ and AUPRC $0.352$ for predicting failure within a $15$ step horizon. It outperforms both a mean difference direction and an action disagreement baseline. A sparse layer sweep finds uneven decodability across depth: layer 16 is strongest among the tested layers, layer 8 remains informative, and layer 10 is weaker. To check whether the monitor is just an occlusion detector, we also evaluate color shift and camera jitter without refitting. Color shift produces no failures in this setting, so it is a benign control rather than a failure benchmark. Camera jitter does induce failures, and the occlusion trained monitor remains above random. The result is deliberately limited: OpenVLA internal states contain failure relevant structure under controlled perceptual shift, but these experiments do not establish a causal mechanism, task held out generalization, or a deployable recovery system.

Vision-language-action models (VLAs) achieve state-of-the-art performance on many robotic manipulation tasks, yet they can still behave unpredictably in out-of-distribution scenarios. Runtime failure detection is therefore essential for the safe real-world deployment of VLAs.However, existing task failure detectors require computationally expensive action sampling, are based on architectural assumptions that limit their applicability to VLAs, or need access to failure rollouts. We propose VLA-FAIL, a lightweight and broadly applicable failure detection framework for VLAs that combines two novel failure detectors with minimal overhead, without requiring failure data. The first, last-layer Mahalanobis distance (LLMD), detects out-of-distribution states by measuring token-wise deviations in last-layer features relative to the training data. The second, action chunk consistency (ACC), exploits the temporal overlap induced by receding-horizon control and detects failures when consecutive action chunks become inconsistent. To capture the trade-off between detection accuracy and detection latency, we introduce AUCPDT, a threshold-independent metric that jointly evaluates precision, recall, and detection time. Through extensive real-world and simulation experiments, we demonstrate that LLMD and ACC capture complementary failure modes whose combination enables reliable and early failure detection across diverse tasks, frequently outperforming significantly more expensive baseline methods. Project page: https://vla-fail.github.io/vla-fail-2026

Vision-Language-Action (VLA) models are increasingly deployed across diverse tasks, yet they remain black boxes whose physical interactions can cause irreversible harm, making generalizable and interpretable failure detection essential. We observe that successful and failed rollouts carry systematically different information-theoretic signatures. Building on this, we formalize VLA control as a closed-loop information pipeline and derive the Triple Information-theoretic (Tri-Info) signals that capture whether actions remain diverse, temporally consistent, and coupled to state transitions. Across six VLA models and three benchmark environments, Tri-Info matches the strongest baselines in-domain. Moreover, Tri-Info transfers across architectures, environments, and the sim-to-real gap without retraining with labeled data, reaching 70% accuracy on real-world tasks. This establishes Tri-Info as a simple yet powerful method that not only detects failures with strong cross-domain generalization, but also delivers interpretable diagnostics of the underlying failure modes.

Vision-Language-Action (VLA) models have shown strong performance in robotic manipulation, but reliable uncertainty quantification remains challenging, particularly under distribution shift. Unlike autoregressive policies, many modern VLA models generate continuous actions through regression or flow-based generation, where explicit predictive probabilities are unavailable. Moreover, stochastic action sampling primarily captures action-generation variability under a fixed model, while failure detection under distribution shift can benefit from capturing uncertainty in the model itself. Motivated by Bayesian perspectives on local model variations, we propose perturbation-based failure detection (PFD), a training-free framework for estimating epistemic uncertainty in VLA models through low-rank weight perturbations. Specifically, we inject random low-rank perturbations into selected transformer weight matrices and estimate epistemic uncertainty from disagreement across perturbed action predictions. Experiments on LIBERO-PRO show that PFD achieves the highest average AUROC and balanced accuracy among the evaluated methods while consistently outperforming stochastic action sampling across distribution shifts. Real-world robot experiments further demonstrate that PFD provides a competitive failure-detection signal under an unseen object shift.

Vision-Language-Action (VLA) models are increasingly used as generalist robot policies, yet their evaluation still relies largely on static benchmarks that randomly sample task scenes. In high-dimensional embodied spaces, failures are sparse and clustered, so static benchmarking can underestimate robustness risks. We reframe VLA evaluation as an active failure-discovery problem and propose a failure-aware test-generation approach that combines diversity-driven exploration with surrogate models learned from observed executions. The method steers testing toward high-risk yet diverse scene regions. Across four state-of-the-art VLA models, it uncovers substantially more failures (up to +29.7 % over selected baselines) while revealing more diverse failure modes. This mean that, for instance, in the case of GR00T-N1.6, success rate dropped from 64.4% to 34.7%. More broadly, our findings call for a shift in VLA evaluation: from passive measurement on fixed task suites to adaptive, failure-seeking test generation that exposes the structure of model weaknesses before deployment.

Vision-Language-Action (VLA) models enable robots to follow natural language instructions and generalize across diverse tasks, but they remain vulnerable to execution failures that compromise reliability in real-world deployment. Detecting such failures during execution is therefore critical for the robust deployment of embodied systems. Existing failure detection methods either rely on expensive action resampling or external models, while alternatives propagate trajectory-level labels uniformly across every timestep, obscuring localized failure signals. In this paper, we propose Hide-and-Seek, a framework that formulates VLA failure detection as a coarsely supervised learning problem. By combining inter-trajectory and intra-trajectory contrastive objectives, Hide-and-Seek localizes failure-indicative actions and induces temporally structured failure signals from trajectory-level supervision alone, without any step-level annotation. We evaluate Hide-and-Seek on LIBERO, VLABench, and a real-world robotic platform across three representative VLA policies: OpenVLA, $π_0$, and $π_{0.5}$.Our method achieves state-of-the-art multi-task failure detection performance with a practical accuracy--timeliness trade-off under conformal prediction, and generalizes well to both seen and unseen tasks.

Vision-Language-Action (VLA) models have demonstrated strong potential for predicting semantic actions in navigation tasks, demonstrating the ability to reason over complex linguistic instructions and visual contexts. However, they are fundamentally hindered by visual-reasoning hallucinations that lead to trajectory deviations. Addressing this issue has conventionally required training external critic modules or relying on complex uncertainty heuristics. In this work, we discover that monitoring a few attention heads within a frozen VLA model can accurately detect path deviations without incurring additional computational overhead. We refer to these heads, which inherently capture the spatiotemporal causality between historical visual sequences and linguistic instructions, as Navigation Heads. Using these heads, we propose an intuitive, training-free anomaly-detection framework that monitors their signals to detect hallucinations in real time. Surprisingly, among over a thousand attention heads, a combination of just three is sufficient to achieve a 44.6 % deviation detection rate with a low false-positive rate of 11.7 %. Furthermore, upon detecting a deviation, we bypass the heavy VLA model and trigger a lightweight Reinforcement Learning (RL) policy to safely execute a shortest-path rollback. By integrating this entire detection-to-recovery pipeline onto a physical robot, we demonstrate its practical robustness. All source code will be publicly available.

Current research on Vision-Language-Action (VLA) models predominantly focuses on enhancing generalization through reasoning techniques. While effective, these improvements increase computational complexity and inference latency. Furthermore, these mechanisms are typically applied indiscriminately, wasting resources on trivial tasks while failing to provide the uncertainty estimation necessary to prevent catastrophic failure on out-of-distribution scenarios. Inspired by human cognition, we propose an adaptive framework that dynamically routes VLA execution based on the complexity of the perceived state. Our approach transforms the VLA's vision-language backbone into an active detection tool by projecting latent embeddings into a set of parametric and non-parametric estimators. This allows the system to execute known tasks immediately (Act), reason about ambiguous scenarios (Think), and preemptively halt execution when encountering physical or semantic anomalies (Abstain). We find that a Gaussian Mixture Model fitted to fused vision-language embeddings provides the most reliable task-complexity signal, combining visual novelty with instruction context and cross-modal compatibility. Evaluated on the LIBERO and LIBERO-PRO benchmarks as well as on a real robot, our fused configuration achieves up to 87.5% F1-score across two VLA backbones (SmolVLA and $π_0$), retains 83% with as little as 5% of training data, and surpasses state-of-the-art failure detectors.

Robust robotic manipulation requires reliable failure detection and recovery. Although recent Vision-Language Models (VLMs) show promise in robot failure detection, their generalization is severely limited by the scarcity and narrow coverage of failure data. To address this bottleneck, we propose an automatic framework for generating diverse robotic planning and execution failures across both simulated and real-world environments. Our approach perturbs successful manipulation trajectories to synthesize failures that reflect realistic failure distributions, and leverages VLMs to produce structured step-by-step reasoning traces. This yields GuardianFail-36k, a large-scale failure reasoning dataset built upon the RLBench simulator and the BridgeDataV2 real-robot dataset. Using GuardianFail-36k, we train Guardian, a multi-view reasoning VLM for unified planning and execution verification. Guardian achieves state-of-the-art performance on three unseen real-world benchmarks: RoboFail, RoboVQA, and our newly introduced UR5-Fail. When integrated with a state-of-the-art LLM-based manipulation policy, it consistently boosts task success rates in both simulation and real-world deployment. These results demonstrate that scaling high-quality failure reasoning data is critical for improving generalization in robotic failure detection. Code, Data, and Models are available at https://www.di.ens.fr/willow/research/guardian/.