acceptodds
Under review as a conference paper at ICLR 2027

LIBERO-Risk: Safety-Critical Scenario Generation and Process-Level Evaluation for Vision-Language-Action Models

Abstract

Vision-Language-Action models (VLAs) map visual observations and language instructions to robot actions. Successful task completion, however, does not necessarily imply safe execution. Evaluating VLA safety therefore requires introducing meaningful risks into source tasks while preserving their semantics and executability. Automating this process is challenging since naive risk injection can produce invalid scenes, alter source-task semantics, or render the task unexecutable. In this paper, we present LIBERO-Risk, the first framework for automated safety-oriented scenario construction in VLA evaluation. Through staged verification and iterative revision, LIBERO-Risk preserves source-task semantics and executability while refining invalid candidates into valid risk scenarios. We further introduce a multidimensional process-level scoring strategy that evaluates VLA safety through contact, spatial, and temporal signals. Experiments show that LIBERO-Risk succeeds in 174 of 234 (74.4%) construction trials and in 64 of 78 (82.1%) task–mutation families. Evaluations across six VLAs further expose safety-relevant behavior that task success alone does not capture.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Vision-Language-Action Models, Safety Evaluation, Scenario Generation, Embodied AI

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_BeeWOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

LIBERO-Risk is proposed as an automated construction pipeline that injects safety constraints into existing manipulation tasks. Compared with existing safety-related benchmarks, the pipeline is fully automated, achieves a construction success rate of roughly 75%, and can be applied to existing tasks while preserving the original task semantics. Its evaluation goes beyond task success to include a process-level safety score.

Critical strengths
  1. Compared with existing safety-related benchmarks, this pipeline is automated, achieves a success rate of approximately 75%, and can be applied to existing tasks while preserving the original task intent. The mutation mechanisms include forbidden-object layout (FOL), physical blocking (PB), and trajectory-bound hazard layout (TBHL).
  2. The evaluation incorporates both task success rate and process-level safety score, deriving five process-safety measurements that characterize contact, spatial proximity, and temporal risk.
Critical weaknesses
  1. The risk types are limited and far removed from real-world usage. Whether such an automated safety pipeline is even needed also remains open to question.
  2. The experiments cover only basic models and do not include recent state-of-the-art models, such as π0.5.
  3. The evaluation is restricted to LIBERO, which is relatively outdated, and thus cannot sufficiently support the robustness of the automated pipeline.
Questions for authors
  1. Consider more useful risk types that reflects real world application.
  2. Evaluation on other dataset except for LIBERO series.
  3. Add more recent models into the experiments.
Reviewer_RHoFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes an automated pipeline for generating safety critical manipulation simulation scenarios by mutating a base scenario. The goal is to provide an evaluation and training tool for VLAs to explore safety. The goal is very timely and interesting, and aims to address a real blind spot in robot training and evaluation benchmarks. The paper proposes a safety scenario generator module, a validity verification module, and a safety scoring module.

Critical strengths

The main strength of this work is the vision to build a safety evaluation simulation benchmark for robot manipulation. This is a very timely and important goal, and to my knowledge no such benchmarks exists to fill this need. The decomposition of the method into the three modules of scenario generation, validity verification, and safety scoring are also a useful and interesting framework for approaching the construction of such a benchmark. The approach is agentic, with a VLM agent programmatically acting to fill the role of each of these modules.

Critical weaknesses

The main weakness of this work is in the execution, richness, and realism of the framework, not the framework itself. I raise my specific concerns in order of decreasing importance:

  1. The running example of a doll in the scene is not very compelling. I would want images and descriptions of non-trivial generated safety scenarios as confidence that the method doesn’t just put dolls in inconvenient locations in every scene. The human quality rating in Fig 4 is very valuable, but also calls into question the quality of the scenario generator. It shows the scenario generator scoring ~3.4 versus a random scene at ~2.7. So it seems like the generator is not building very compelling scenarios as judged by humans. In general, the paper would benefit from more visual examples of these safety scenarios to the readers can judge how rich and realistic they are.

  2. The Scenario mutation module only uses the objects in the original task set? If so, this seems problematic. A key attribute of safety critical situations is that they often involve unexpected objects and interactions. The scenario generator should aim to construct such interactions by introducing new objects or scene interactions not present in the original task. Please clarify how this works.

  3. How do we validate the validator? Validity is subject to the quality of the VLM that does the assessment. Does the VLM assess validity through visual feedback from the simulator, or does it have simulator state information? If it has state information, how would this transfer to real world robot safety assessment? Executability in particular seems problematic. The text says “replay the reference execution \tau_ref to ensure that the mutation does not render the task infeasible and that at least one successful execution path remains.” This makes me question where these executions come from. Does the VLA have to already generate a safe trajectory? Or who generates the safe trajectory?

  4. The safety scoring module is more quantitative than the validity module, but subjective metrics like “how close is too close” remain problematic for this module.

  5. Where is the technical novelty in the method itself? It is all based on VLM prompting, I believe?

  6. Safety is often context driven. E.g., pouring water on a laptop is a safety violation usually, unless the task is robustness testing of the laptop against water hazards. Sometimes the same object can be benign or a critical safety risk depending on the user, the task, and the context of the task. Safety is very complex. How does your method address the context driven nature of safety?

  7. I would argue that real safety violations often occur through cascading faults in the robot system, not because of objects that are inherently unsafe. For example, an object slip or bump during an attempted grasp excites a low-level control oscillation, that adversely feeds back into the VLA, which lead to a wild arm swing that breaks objects in the scene. I would argue this is the honest source of safety problems for VLAs. I think this method does not address these kinds of real-world system interaction and integration issues. Can you comment on how your method might include such system-level problems?

Questions for authors

My questions are embedded together with my comments above.

1 comment

I appreciate the authors' detailed response. I understand this takes quite a bit of effort. Unfortunately, I am going to keep my score as is. Two hurdles are stopping me from raising: 1. If pure collision avoidance is the primary concern, there are quite strong classical methods to accomplish this, e.g. a control barrier function as a safety layer modifying the VLA output (operating on a mono-to-depth model from the image). If it is just collision avoidance, there is no need for the semantic nature of the benchmark. 2. The engineer deep inside me says this is not why robot's fail. They fail due to cascading system level errors and interactions that cannot be replicated in a simulation benchmark with a closed object set. It just not rich enough to find to real system faults. I applaud the research direction and wish the author team the best of luck!

Reviewer_sU7BOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LIBERO-Risk constructs safety-oriented mutations of existing manipulation tasks using an agent, staged validity checks, and iterative repair. It reports contact, clearance, and temporal risk separately from task success. Experiments cover 234 construction trials, ablate generation and verification, and evaluate six VLAs on 28 source-mutation pairs, with complementary human judgments.

Critical strengths

S1: Verification feedback improves construction reliability. Table 1 reports 73.3% construction success for the complete pipeline versus 46.1% without feedback. Applying the same final verifier across settings makes this comparison informative about repair rather than merely changing acceptance criteria.

S2: Matched evaluations separate task completion from process risk. Section 4.3 evaluates six policies under paired initializations and reports success, safety, and efficiency separately. Successful rollouts can still exhibit proximity or temporal risk, making the evaluation more informative than success alone. Figure 4 reports higher perceived risk exposure after mutation and includes uncertainty. Section 4.4 also acknowledges that the contact and unsafe-event increases are inconclusive, appropriately distinguishing these outcomes from the exposure result.

Critical weaknesses

W1: Avoidable policy risk is not separated from scenario difficulty. Section 3.2 verifies the stated task-executability requirement by successful reference replay, not safety-constrained success. This supports construction validity, but does not determine whether violations in Section 4.3 are avoidable policy mistakes or imposed by hazard placement. Show safe successful references for a representative sample, or explicitly limit policy-level attribution where safe completion is unestablished.

W2: Descriptive risk conclusions lack sensitivity analysis. The main text lists five process-safety measurements but does not clearly state how they are aggregated or how the risk thresholds are chosen. Figure 4 supports increased human-observed exposure after mutation, but it does not report per-rollout agreement between the binary indicators and human labels. Report indicator agreement on the rated rollouts, test threshold sensitivity of the 96.4% risk prevalence, and assess the stability of Table 2's aggregate policy-score comparisons. These checks matter because the component measurements, thresholded prevalence, and weighted policy scores support different conclusions.

Questions for authors
  1. Are safe successful references available for a representative sample, and which policy-risk conclusions depend on assuming avoidability?

  2. How do automatic binary risk indicators agree with human labels on the same rated rollouts?

  3. How do plausible threshold or weight changes affect the 96.4% prevalence and Table 2's policy comparisons, separately?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.