- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper investigates whether relationships among internal signals in vision-language-action (VLA) models can predict eventual task failure more accurately and earlier than existing failure predictors. The method collects attention weights, value vectors, hidden states, embedding gradients, observations, and actions from VLA inference. Handcrafted statistics such as cosine similarity, Spearman correlation, top-k IoU, and Wasserstein distance characterize relationships between these signals.
Critical strengths
-
Important and practically motivated problem. Predicting failure before a robot irreversibly executes an unsuccessful trajectory is an important challenge.
-
Comprehensive experiments. The inclusion of autoregressive and continuous-action VLA architectures increases the relevance of the study. It also highlights that different architectures expose different internal signals. Meanwhile, this paper also has a detailed feature analysis.
Critical weaknesses
-
The approach primarily constructs manually designed features from existing internal signals and trains a GRU classifier. It is unclear whether the proposed relationships are fundamentally better than appropriately trained predictors using the same underlying signals.
-
The experiments do not isolate the source of improvement. Better performance might come from richer input features, temporal modeling, feature selection, or different training procedures rather than the proposed introspective relationships.
-
The practical claim of early failure prediction is insufficiently demonstrated. Per-step AUROC improvements do not necessarily imply useful early warnings, especially without reporting detection lead time, false-alarm tradeoffs, and inference overhead.
Questions for authors
-
How much of the proposed method's improvement remains when all internal attention, hidden-state, and gradient-based features are removed, retaining only observation and action-consistency features?
-
Can a GRU trained on the same underlying signals, without handcrafted pairwise relationships, achieve comparable accuracy?
-
How sensitive is the performance to the action-related scalar used to compute embedding gradients?
3 comments
Response to Reviewer jcoc (2/3)
The practical claim of early failure prediction is insufficiently demonstrated. Per-step AUROC improvements do not necessarily imply useful early warnings, especially without reporting detection lead time, false-alarm tradeoffs, and inference overhead.
We address this concern in two parts:
(1) Lack of Ground Truth for Failure Onset. Among the four baselines (i.e. SAFE [1], FIPER [2], INSIGHT [3], FAIL-Detect [4]), only SAFE compares detection timing with human-annotated failure times (INSIGHT only annotates steps that don’t make progress towards the goal as a help trigger signal). However, SAFE uses judgments of when intervention is needed for ambiguous cases, not of when a failure clearly onsets. Since there is no gold standard or ground truth to identify when failure happens, we attempted at having a human annotator (one of the authors) and an LLM (GPT 6.1-sol) independently annotate 50 failed rollouts from each of the three policies. The audit revealed substantial disagreement in the failure onset steps between the human annotator and the LLM, with Mean Absolute Differences (MAE) of 38.1, 32.6, and 30.7 policy steps for OpenVLA-OFT, π0.5, and π0-FAST, respectively. This disagreement suggests that judgement of failure onset is ambiguous and difficult to scale for large evaluations. For example, one annotator may rule the step where the robot mis-grasps and drops an object as failure onset, whereas another may wait until the robot reaches an unrecoverable failure state. SAFE also discusses similar ambiguity in their human annotation process in Appendix C.3.
(2) New Analysis: Normalized First-Detection Time in Failure Episodes. We aim to predict eventual failures as early as possible within the rollout while using the final episode outcome as the supervision and evaluation target. Following the rollout-relative timing evaluations in FIPER and SAFE, we define normalized first-detection time as the first-detection step divided by episode length. Following FIPER, we compute it over correctly detected failure episodes. We report results averaged over 40 folds at thresholds calibrated to target episode-level FPRs of 5%, 10%, and 20%. Baselines are evaluated at their original operating points. Lower values indicate earlier detection. At the 10% FPR target, our predictor first detects failure after 13.1%, 13.0%, and 13.6% of the rollout for , -FAST, and OpenVLA-OFT, respectively. These detection times are earlier than those of every applicable baseline at its original operating point. We have included this analysis in more detail in the updated manuscript Analysis section.
| VLA model | Ours 5% FPR | Ours 10% FPR | Ours 20% FPR | SAFE | FAIL-Detect | INSIGHT | FIPER |
|---|---|---|---|---|---|---|---|
| 0.169 ± 0.043 | 0.131 ± 0.038 | 0.089 ± 0.032 | 0.214 ± 0.142 | 0.383 ± 0.176 | — | 0.476 ± 0.104 | |
| -FAST | 0.131 ± 0.050 | 0.130 ± 0.045 | 0.118 ± 0.048 | 0.164 ± 0.149 | 0.276 ± 0.172 | 0.291 ± 0.142 | — |
| OpenVLA-OFT | 0.178 ± 0.042 | 0.136 ± 0.035 | 0.106 ± 0.029 | 0.169 ± 0.145 | 0.232 ± 0.100 | — | — |
[1] Gu, Qiao, et al. "Safe: Multitask failure detection for vision-language-action models." Advances in Neural Information Processing Systems 38 (2026): 40041-40076.
[2] Römer, Ralf, et al. "Failure prediction at runtime for generative robot policies." Advances in Neural Information Processing Systems 38 (2026): 7631-7670.
[3] Karli, Ulas Berk, et al. "Insight: Inference-time sequence introspection for generating help triggers in vision-language-action models." 2026 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2026.
[4] Xu, Chen, et al. "Can we detect failures without failure data? uncertainty-aware runtime failure detection for imitation learning policies." arXiv preprint arXiv:2503.08558 (2025).
Response to Reviewer jcoc (3/3)
How much of the proposed method's improvement remains when all internal attention, hidden-state, and gradient-based features are removed, retaining only observation and action-consistency features?
We provide a comparison between the compact feature set predictor and a predictor trained on only observation and action-consistency features for OpenVLA-OFT, π0.5, and π0-FAST, respectively. When all internal attention, hidden-state, and gradient-based features are removed, the test AUROCs worsen drastically at step 0-5 across all three policies and remain relatively inferior at step 5-15 for π0.5 and OpenVLA-OFT. This suggests the features constructed on the internal signals are very helpful at the early phase of policy rollouts. Please see a figure of per-step task-disjoint test AUROC comparison in Section A.4 in the updated manuscript.
How sensitive is the performance to the action-related scalar used to compute embedding gradients?
We'd like to clarify that the action-related scalar is used only to compute gradient-based attribution; it is not provided directly to either the compact raw signal set predictor or the compact feature set predictor. As reported in Appendix A.1 of the updated manuscript, gradient-based attribution features are frequently useful: several rank among the top 30 single-feature GRUs and enter the starter pool for forward selection, and some are selected in the final six-feature sets. However, the predictor does not depend on them. The fixed six-feature GRUs for and -FAST used in our main comparison to baselines contain no gradient-derived features and still achieve similar per-step AUROC. Thus, gradient-based attribution provides useful complementary evidence when selected, while compact predictors can also perform well without it. Our current experiments establish the contribution of gradient-based attribution features but do not compare alternative definitions of the action-related scalar directly.
Response to Reviewer jcoc (1/3)
Thank you for taking the time to review and providing suggestions to strengthen our analysis. We would like to address your concerns and questions as follows:
Thank you for suggesting this analysis. We address these questions in two parts:
(1) New Analysis: Very Limited Gain from Richer Inputs. We trained a compact raw signal predictor using six inputs selected from attention weights, value vectors, hidden states, gradient-based attribution signals, and action outputs, etc. The predictor uses the same GRU architecture and training procedure as our relational feature predictor. Please see a more detailed training description in Section 6 of the updated manuscript. We compare it with SAFE, which operates on policy hidden states only. The broader raw signal pool provides a modest improvement for π0.5 and similar performance for π0-FAST and OpenVLA-OFT. Thus, access to richer raw inputs alone produces limited gains and does not explain the improvement from our relational feature construction. Please see a figure of per-step task-disjoint test AUROC for SAFE and the compact raw signal set predictor in Section A.3.
New Analysis: Failure Prediction from Raw Signals vs. Relational Features To address whether improvements come from relational feature design and temporal modeling, we also provide comparison between the compact raw signal set predictor and a compact feature set predictor (Training detail described in Section 6.1 of the updated manuscript). The relational feature set substantially outperforms the raw signals throughout the rollout for π0.5 and π0-FAST. For OpenVLA-OFT, the two predictors perform similarly during the first 30 steps, after which the relational feature set is generally stronger, with some variation near the end of the rollout. These results show that explicitly modeling relationships among internal signals provides predictive information beyond the raw quantities and their direct temporal changes. We have included this analysis in our updated manuscript Section 6.3. Please also see a figure of per-step task-disjoint test AUROC for compact raw signal and relational feature sets in Section 6.3.