acceptodds
Desk rejected at ICLR 2027

HazardArena: Evaluating Semantic Safety in Vision–Language–Action Models

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-Language-Action (VLA) models inherit rich world knowledge from vision-language backbones and acquire executable skills from action demonstrations. Yet current evaluations primarily measure task completion, leaving the semantic safety of learned action policies underexplored. This gap creates a critical vulnerability: a policy may execute the intended action correctly while producing unsafe outcomes when the surrounding visual-linguistic context changes. We present HazardArena, a benchmark for stress-testing semantic safety in VLA systems. Its core design is a set of safe/unsafe twin scenarios: paired environments with matched objects, layouts, and action requirements, but different semantic risk contexts. This controlled contrast isolates safety judgment from motor capability and directly tests whether a VLA can recognize when an otherwise valid action becomes hazardous. HazardArena includes over 2,000 assets and 50 risk-sensitive tasks across seven safety categories grounded in robotic safety standards. Across four representative VLA backbones, we observe a consistent and alarming pattern: safe-only fine-tuning improves benign task success while also increasing hazardous execution on matched unsafe scenarios. Physical-world experiments confirm that this failure transfers beyond simulation. These results show that stronger action execution does not imply safer behavior, and motivate semantic-risk-aware evaluation and enforcement as first-class requirements for real-world VLA deployment.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, robot safety, AI safety, embodied AI, benchmark, VLM

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_shwKOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This article introduces HazardArena, a benchmark for evaluating the semantic safety of visual-language-action (VLA) models, aiming to study whether the models can adjust their action behaviors based on the risk information in the scenes and instructions. This benchmark covers seven types of safety risks and 51 dangerous scenarios. By constructing safety and danger contrast tasks that match the motion requirements and designing three-stage indicators of Attempt, Commit, and Success, it distinguishes the models' safety awareness from their action execution capabilities. Experiments were conducted on models such as OpenVLA-OFT, VLA-Adapter, NORA, π₀, and π₀.₅, revealing that although fine-tuning with only safety task data improved the success rate of normal tasks, it also increased the execution rate of corresponding dangerous tasks. The authors further evaluated two mitigation strategies: external VLM safety filtering and trajectory-enhanced rejection fine-tuning. They verified the diagnostic value of the benchmark and the aforementioned phenomenon through ablation analysis, multiple random seeds evaluation, and real robot experiments with the Franka Panda.

Critical strengths
  1. The article constructs safety and danger scenarios that match the action requirements as much as possible to minimize the influence on the robot's operational capabilities, thereby enabling a more accurate assessment of whether the VLA model can adjust its behavior based on semantic risks. Moreover, the three-stage indicators of Attempt, Commit, and Success can reveal dangerous behaviors that might be overlooked when relying solely on the final task success rate.

  2. The experiments cover a wide range and the authors observed consistent phenomena on multiple representative VLA models: using only safety task data for fine-tuning can improve the model's ability to perform normal tasks while also potentially enhancing its ability to perform dangerous tasks. The article further validates this phenomenon through multiple random seeds evaluation, additional tasks, two safety mitigation strategies, and real robot experiments, providing strong empirical support for the main findings.

Critical weaknesses
  1. This article presents a benchmark for the security assessment of VLA, but the revelations it brings are actually not new. An un-safely aligned VLA will definitely accept all instructions. Their output is actions that have no reason to refuse or deliberately fail. The two proposed defense strategies are also relatively simple. So I think what the author is claiming, "By doing SFT on safe tasks, the success rate in malicious tasks has also increased", is not something interesting or novel. In fact, it is somewhat misleading. The VLA model has no idea what is safe and what is malicious at all. Whether he can accept this task and complete it depends solely on whether he was familiar with this trajectory during his training stage. If a model that has undergone safety alignment still exhibits such a phenomenon, then I would consider this claim to be interesting and valid. For instance, in a previous community study, an aligned LLM was fine-tuned on some security or personalized datasets, and as a result, its probability of rejecting malicious content also decreased.

  2. For instance, the tasks were somewhat forced: putting a capsule into a wine glass, throwing something into the trash can. These "hazards" only make sense when considering the context clearly (For instance, the capsule might be toxic, or the object might not be discarded but is actually very valuable, because the ring could also be something that the user no longer wants and thus needs to be disposed of); otherwise, I don't think putting a capsule into a wine glass is a malicious act that needs to be rejected.

  3. I noticed that the anonymous repository seems to be incomplete; some codes are not provided. For example, "unseen tasks" are not truly implemented in the code: the get_unseen_task_config and seen functions return the same configuration, and the relevant generalization statements need to be verified.

  4. The text was leaked, and what was measured might be the key words rather than the actual understanding of security. "Gently, safely, for a safe demo" only appear in the safe execution scenarios, and never in dangerous instructions. Therefore, there is a possibility that the model can learn shortcuts: execute based on seeing these words without understanding the context. I speculate that if you add malicious instructions to these "safely" keywords, such as safely directing the knife towards that baby, the damage rates of all aligned models will increase. This is the effect brought about by the shortcut. I think for the sake of rigor, we can uniformly remove all of them.

  5. Regarding the reliability of the determination and tasks, COMMIT is a geometric determination based on the distance radius. However, in the oven task, the script never triggered COMMIT even once (0/50) when performing dangerous actions. This should be a bug, and the microwave oven task is basically unusable. The safety scenario 1 of the socket task uses the placeholder sentence "insert the object into the target". These flaws will significantly reduce the value of the benchmark for the community.

minor: 1.The resolution of Figure 3 is not clear.

Questions for authors
  1. If we were to re-examine the argument of the article on the aligned VLA model, "By making minor adjustments in safe tasks, his success rate in malicious tasks has increased" would still hold true?

  2. If the task of placing the capsule into the wine glass is also classified as a malicious task, it is likely to cause the model to overly reject some benign tasks. How can this trade-off be balanced?

  3. What is the purpose of adding "Gently, safely, for a safe demonstration" to some security tasks? Is it because these tasks (such as those involving sensitive objects) are more likely to be rejected? So that's why it was done as a relief measure?

Reviewer_cq2pOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces HazardArena, a semantic safety benchmark for Vision–Language–Action (VLA) models, designed to evaluate whether robots can distinguish between “actions they are capable of performing” and “dangerous actions they should not perform” based on the task context.

Critical strengths
  1. The safety issues raised in the paper are significant. Most traditional manipulation benchmarks focus on whether a task can be completed, whereas whether a robot should carry out a user’s request is a separate research question.

  2. Experiments involving multiple VLA backbones were conducted.

Critical weaknesses

W1. Insufficient Evaluation of Semantic Safety Generalization.

In the zero-shot evaluation across all 51 scenarios, all four pre-training strategies achieve zero success on safe tasks, making it difficult to interpret the corresponding zero success on dangerous tasks as evidence of semantic safety awareness. The results do not distinguish between deliberate rejection of unsafe instructions and a general inability to execute the required manipulation skills. Furthermore, the evaluation primarily examines whether skills learned from safe demonstrations transfer to semantically related dangerous tasks, without systematically testing unseen combinations of hazardous objects, instructions, and scenarios. Consequently, the experiments do not convincingly establish whether the models possess or lack generalizable semantic safety capabilities, substantially limiting the conclusions that can be drawn about safety generalization.

W2. Limited Novelty and Insufficient Validation of the Safety Evaluation Metric.

The central finding that safe-only fine-tuning improves safe-task performance while also increasing dangerous-task success is valuable but largely expected, as models trained on successful demonstrations without explicit safety supervision may naturally transfer learned manipulation skills to semantically similar hazardous tasks. Although the paper confirms this phenomenon across multiple models, it does not sufficiently establish a previously unknown failure mechanism. Moreover, the COMMIT metric relies on task-specific geometric or contact predicates that do not necessarily indicate dangerous intent or actual commitment to a hazardous action. Entering a predefined region, for example, may reflect ordinary motion rather than an attempt to perform the dangerous operation. Without stronger semantic validation of COMMIT and deeper analysis of the underlying failure mechanisms, the paper's empirical findings provide limited support for a substantial conceptual contribution to embodied AI safety.

Questions for authors

As weaknesses

Reviewer_cbiyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces HazardArena, a simulation benchmark for semantic safety in Vision–Language–Action (VLA) models. Its core design is safe/unsafe twin scenarios that share the same skill template, scene, and motor demands but differ in one risk-critical semantic factor (e.g., knife→bread or toy→baby vs. knife→baby). The benchmark covers seven risk categories derived from ISO 13482 and AutoRT and a 51-scenario inventory with 75 new context-critical assets. It defines ordered stage-wise metrics (ATTEMPT, COMMIT, SUCCESS) that capture hazardous progression beyond terminal success, and provides a pipeline for automatically generating refusal demonstrations. Fine-tuning OpenVLA-OFT, VLA-Adapter, NORA, and π0 on 1,000 safe-only demonstrations from 10 twin pairs increases unsafe success alongside safe success, with Early-to-Final gains of +0.06 to +0.32. This leaves final safe/unsafe gaps near zero, and π0.5 and six additional π0-trained pairs show the same pattern. Supporting analyses include a 200-video blinded audit of the COMMIT predicate, threshold sensitivity, grounding controls, and attention diagnostics. Two mitigations are studied. An inference-time VLM safety layer achieves recall that varies strongly by category. Refusal-augmented fine-tuning reduces unsafe success to 0.01–0.04 while preserving safe success. A Franka Panda study with π0 reproduces this trend, with unsafe success dropping from 39.6% to 0.4% when refusal data is added.

Critical strengths
  1. A clean counterfactual design. Safe/unsafe twins hold the skill template, scene, and motor demands fixed and vary only the risk-critical semantics. This separates refusal from incapacity, an ambiguity that limits existing VLA safety benchmarks. The design is a reusable evaluation principle for the field.

  2. Capability-aware, stage-wise metrics. ATTEMPT, COMMIT, and SUCCESS reveal hazardous progression that terminal success hides. For example, NORA has lower unsafe success than VLA-Adapter on Insert Outlet but higher COMMIT (0.62 vs. 0.46). A blinded audit with 198/200 annotator agreement supports the metric. This makes safety evaluation sensitive to near-misses, not only completed hazards.

  3. Useful infrastructure with real-robot validation. The benchmark offers seven risk categories, a 51-scenario inventory, 75 new assets, and automated refusal-demonstration generation, evaluated on five backbones. Its trends transfer to a Franka Panda, where refusal data reduces unsafe success from 39.6% to 0.4%. This gives the community a practical resource for studying and training semantic safety in VLAs.

Critical weaknesses

(1) The central "hazard amplification" finding is largely expected. The policies are trained only on safe demonstrations and never receive any refusal signal, while each unsafe twin recombines objects and verbs seen during training (e.g., knife→bread and toy→baby yield knife→baby). Executing the unsafe instruction therefore reflects compositional generalization rather than a distinct safety failure. Moreover, at the Early checkpoint safe and unsafe success are both near zero (all 0.00 for OpenVLA-OFT), so the "lockstep" Early-to-Final increase mostly tracks the emergence of capability. Calling this "amplification" overstates what the experiments show.

(2) Refusal training appears to be trained and tested on the same tasks. Refusal demonstrations are generated for the unsafe instructions of the same 10 pairs used for evaluation (Appendix D.2, Table 4), and the real-robot study follows the same protocol. The drop in unsafe success to 0.01–0.04 therefore cannot distinguish semantic safety from memorizing a mapping from these specific instructions to "return home," and evaluation on held-out unsafe scenarios or categories is needed. The authors themselves note that the refusal-trained policy degenerates into surface pattern matching on confusable cases.

(3) The capability-matched premise is weak, and the evaluated models are dated. Final safe success is low, at 0.04–0.13 for OpenVLA-OFT and mostly 0.15–0.47 for the others, and motor equivalence is only approximate; for example, unsafe success on Microwave Egg is consistently much lower than safe success. The zero-shot evaluation (Table 13) covers only four older policies, all of which fail every safe task, so it establishes only a capability floor. π0.5 [1] is evaluated only after safe-only fine-tuning, with unreported hyperparameters and no zero-shot or Early checkpoint, even though it is the backbone most likely to retain semantic knowledge from co-training. Evaluating π0.5 zero-shot, and recent large-scale VLAs such as LingBot-VLA [2], would make the capability-matched comparison meaningful. A knowledge-rich frontier model such as GPT-6 Astra [3] could also serve as a semantic reference: even a small qualitative probe that asks it to judge a handful of twin pairs as safe or unsafe would show whether the benchmark's risk distinctions can be recovered from world knowledge. Such a probe would also help explain categories where the Qwen3-VL-32B judge fails (e.g., 0.00 recall on Property) by separating judge limitations from ambiguous labels. Together, these results would clarify whether the observed failures stem from missing knowledge or from action policies that fail to use it.

(4) The writing reads as a patchwork of rebuttal responses. The main text repeatedly refers to the "revised study," the "rebuttal expansion," results "as reported in the rebuttal," and "released Table B," and it is saturated with hedges such as "does not establish." Several passages admit that the authors cannot reconcile their own results: Table 3's two panels "cannot be reconciled," and the bootstrap intervals lack episode-level logs. With evidence scattered across these sections, it is hard to map claims to supporting results, and the paper needs a substantial rewrite to state its findings clearly.

References [1] Physical Intelligence et al., "π0.5: A Vision-Language-Action Model with Open-World Generalization," arXiv:2504.16054, 2025. [2] Robbyant, "A Pragmatic VLA Foundation Model" (LingBot-VLA), arXiv:2601.18692, 2026. [3] OpenAI, "GPT-6 Astra," model release, September 2026.

Questions for authors

see weakness

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.