acceptodds
Under review as a conference paper at ICLR 2027

ASR-VLA: Adaptive Safety Reasoning with Safety-Critical Future Prediction for Vision-Language-Action Models

Abstract

Vision-Language-Action (VLA) models have demonstrated strong capabilities in language-conditioned robotic manipulation. However, safety remains a critical and insufficiently addressed challenge for VLA models. Existing safety-aware VLA methods primarily formulate safety as a policy constraint or apply external correction to potentially unsafe actions. They generally do not explicitly ground action generation in a predicted, state-dependent future hazard and reason about how the robot should respond. Consequently, the policy may struggle to determine an appropriate response to evolving risks while maintaining task progress. To address this problem, we introduce ASR-VLA, an adaptive framework that integrates explicit safety reasoning into VLA policies. ASR-VLA couples a risk-aware world model with a vision-language reasoning model. To support this reasoning, ASR-VLA incorporates a risk-aware world model that predicts the most safety-critical future observation as auxiliary visual evidence. The vision-language reasoning model analyzes the current scene and task context, using the predicted future risk to generate structured safety guidance. The resulting reasoning and future visual evidence jointly condition action generation, enabling the policy to adapt its behavior while maintaining task progress. Experiments on SafeLIBERO demonstrate that ASR-VLA substantially improves both task performance and safety, achieving a 93.68% success rate with only a 9.46% collision rate. Compared with the strongest baseline, ASR-VLA increases task success by 20.72% while reducing collisions by 10.46%, yielding a markedly better success–safety performance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Reasoning Module.

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_XGRdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work use a world model to predicts the most dangerous future scene, then uses a VLM to reason about safety guidance, and feeds both the predicted image and guidance into a VLA to help the robot avoid obstacles while completing the task.

Critical strengths
  1. The paper is well motivated and targets at a critical problem: improving safety in VLA-based robotic manipulation.

  2. The results show improvements in task success and collision avoidance on SafeLIBERO.

  3. The paper includes real-robot demonstrations that provide initial evidence of practical applicability.

Critical weaknesses

Despite some promising results, the experimental design and comparisons have several flaws that limit support for the paper’s main conclusions. The evaluation is limited in benchmark coverage, and the ablation studies do not adequately justify the design choices.

  1. The effect of the training data is not isolated. I could not find a pi0.5 baseline directly fine-tuned on the same SafeLIBERO demonstrations as ASR-VLA. This is missing of experiment is fatal. Comparing against a pi05 policy trained on regular LIBERO is not a fair comparison. Consequently, it remains unclear whether the improvements come from the proposed design or simply from learning obstacle-avoidance behaviors present in the training data.

  2. Limited baseline comparisons. The paper evaluates only a narrow set of baselines and does not clearly show whether they receive the same SafeLIBERO training data as ASR-VLA, making it difficult to conclude the source of the gain. Moreover, neither safety-aware baseline is world-model-based. The absence of a comparison with closely related world-model-based methods, such as SafeDojo, leaves the contribution relative to existing approaches insufficiently established.

  3. Limited evaluation scope. The evaluation relies on a single simulation benchmark with a limited set of tasks focused on tabletop obstacle avoidance. This narrow coverage does not sufficiently support the paper’s broader claims about safety-aware robotic manipulation. As is illustrated in many other research work, the conclusion can be very different between mobile and table-top settings

  4. Limited real-robot evaluation. The real-robot evaluation covers only a few tasks and reports neither trial counts nor success and collision rates, making it difficult to assess whether the demonstrated benefits are statistically meaningful. Moreover, the experiment scenes looks uncluttered and requires minimal collision awareness.

Questions for authors
  1. Can the authors report results for pi05 and OpenVLA directly fine-tuned on the same SafeLIBERO demonstrations as ASR-VLA, without future prediction or safety reasoning, to isolate the effect of the training data?

  2. Can the authors evaluate the framework with a broader range of VLA backbones, such as X-VLA[1] and UniVLA[2], to assess whether its benefits generalize beyond pi05 and open-vla?

  3. Can the authors include world-model-based safe-control baselines, such as SafeDojo, to establish the contribution relative to existing predictive safety approaches?

  4. Both ablated variants achieve strong and similar performance. Can the authors explain why? This seems contradictory to claimed complementary roles of future imagery and safety reasoning. Does the incremental improvement from combining them justify the additional model training and inference costs?

  5. Can the authors provide out-of-distribution evaluation by withholding some task types or an entire suite during training and evaluating on the held-out tasks? The current evaluation does not clearly establish whether the method learns transferable safety behaviors or overfits to the training task distribution.

[1] Jinliang Zheng et al., X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model

[2] Yuqi Wang et al., Unified Vision-Language-Action Model

Reviewer_4GefOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

ASR-VLA emphasizes safety in robotic manipulation, which is very important. However, its design adds a separate reasoning model and a world model to an existing VLA model. This makes the pipeline heavy and runs counter to the goal of improving the existing VLA model to provide these capabilities itself.

Critical strengths
  1. Safety and action feasibility are key issues for VLA models. The paper focuses on these issues and proposes a solution that achieves good performance.

  2. The writing and figures are clear and easy to follow.

Critical weaknesses
  1. A VLA model is already a foundation model and should have substantial knowledge and the ability to predict future outcomes. The main focus should therefore be on improving its own reasoning and prediction capabilities. Instead, the paper adds two separate models: a reasoning model and a world model. This makes the overall pipeline heavy. Comparisons with unified models in both performance and efficiency are needed to justify these extra modules.

  2. In Table 1, adding either Safety Reasoning or Risky-Future Image alone already provides a large improvement over the base model. The information provided by these two components also overlaps. Combining them brings only a small additional gain. Is this small gain worth the extra latency and computational cost?

  3. The reported pipeline latency is about 3.5 times that of the base model. Can this latency support real-time execution, especially if the method is later extended to dynamic and complex scenarios?

Questions for authors

Please refer to the weaknesses section. I am willing to raise the score if the authors can address my concerns.

Reviewer_CSjUOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes ASR-VLA, which adds two modules to a π0.5 policy. (i) A world model, trained on closed-loop π0.5 rollouts, predicts the highest-risk future frame within a 20-step horizon and a clearance-based risk score. (ii) A LoRA-tuned Qwen3-VL-8B turns the current frame, the predicted frame, the instruction, and the risk score into a structured text (task, subtask, risk status, guidance); it is re-invoked only when the risk status changes. The policy is conditioned on the predicted frame and this text and trained with flow matching. On 25 SafeLIBERO scenarios, the paper reports 93.68% SR / 9.46% CR versus 72.96% / 19.92% for the strongest baseline (VLSA), supported by ablations, reasoning-intervention studies, and a qualitative real-robot demonstration.

Critical strengths
  • The paper addresses a relevant gap: instead of only detecting or filtering unsafe actions, it conditions action generation on anticipated hazards and the task context.
  • The risk-selected future target (the highest-risk frame within a horizon, defined from clearance) is simple and interpretable. The structured output and event-triggered invocation are a practical attempt to keep reasoning latency manageable.
  • The paper includes the right kinds of experiments to probe the mechanism: reasoning interventions, temporal-alignment controls, and anticipation analysis. It also reports per-scenario results over five seeds, world-model comparisons against simple baselines, a latency breakdown, and a hardware demonstration.
  • The method is clearly presented, and the appendix documents prompts, templates, hyperparameters, and algorithms in detail.
Critical weaknesses

W1. The baseline results do not follow the stated evaluation protocol. The paper states that all methods are evaluated with 50 configurations × 5 seeds (L364–365; Table 2: 250 rollouts per scenario) and that trainable baselines use the same demonstrations (L345). However:

  • The OpenVLA-OFT and VLSA rows of Table 1 coincide exactly, for all four suites, with the per-scenario results of the VLSA paper (Hu et al., 2025, Appendix Tables 2–5; a single run of 50 episodes per scenario) averaged over the 25 scenarios used here. For example, for OpenVLA-OFT on Spatial, SR = (16+76+20+26+74+6)/6 = 36.33 and CR = 100 − (0+38+0+0+4+32)/6 = 87.67, as in Table 1.
  • None of the eight π0.5 suite-level values can arise from the stated rollout counts. For example, Long has 1,000 rollouts, so its rates must be multiples of 0.1%, yet 34.73/85.89 are reported, whereas all other methods report multiples of 0.1%. Similarly, in the Object column (2,000 rollouts), every method except π0.5 reports multiples of 0.05%.

The comparison in Table 1, and the headline improvements derived from it, are therefore not obtained under a common protocol. This also suggests that results from the VLSA paper were reused directly.

W2. The gains over the baselines cannot be attributed to the proposed components. The training conditions differ across methods:

  • π0.5 and OpenVLA-OFT are trained only on the original LIBERO data, which contain no added obstacles (L338–341).
  • VLSA is a training-free layer on top of such a policy.
  • SafeVLA's training setup is not described.
  • ASR-VLA, in contrast, is fine-tuned on data collected in the obstacle-augmented SafeLIBERO environments, starting from a Task+Subtask checkpoint (Table 6).

Within the method, removing the safety reasoning still yields 91.74% SR. This variant thus accounts for about 90% of the 20.72-point SR gain over VLSA and about half of the 10.46-point CR reduction, while the reasoning itself adds only 1.9 SR points and lowers CR by 5.3 points. The contribution of the risky-future image over plain fine-tuning is unknown, because no variant removes both components. The Task-only and Task+Subtask checkpoints are trained (App. D.1) but not reported.

W3. The paper does not specify how the policy learns to follow the guidance, and the training data may contradict it. The descriptions of the training data conflict:

  • §3.2 says the world-model data are π0.5 rollouts containing near-collision and collision states.
  • §3.4 calls the policy targets "expert action chunks".
  • §4.1 uses one dataset for all three modules.
  • App. E (L1245–1253) states that, "following the same procedure used in SafeLIBERO", nominal-policy rollouts with "safe, low-clearance, near-collision, and collision states" train "the world model, reasoning model, and safety-grounded VLA policy".

The consequences depend on which data supply the action labels:

  • Labels from nominal rollouts (as App. E states). The risk status is computed from the future of the same trajectory (Eqs. 6–7, 13), or predicted by a world model trained on that trajectory. DANGER is therefore attached exactly to actions that lead into low clearance, and the unweighted flow-matching loss (Eq. 19) trains the policy to execute those actions while receiving a DANGER warning.
  • Labels from safe demonstrations only. The policy imitates safe behavior whatever the guidance says. The safety gain may then come from the demonstrations rather than from risk prediction or reasoning (cf. W2).
  • Mixed labels. The same DANGER input is paired with both risky and safe actions, and nothing in the objective favors the safe ones.

In none of these cases does the objective teach the policy to act on the guidance. That would require pairing DANGER inputs with corrective actions from the same states, or filtering or reweighting the risky samples. The paper describes neither.

W4. The policy sees different inputs during training and deployment, and the ablations do not state what was removed.

  • What the ablations remove, and how, is unclear. "w/o Risky-Future Image" only stops feeding the predicted image to the policy. The reasoning model still looks at that image when writing RtR_t, so the image's information still reaches the policy through the text. Yet the paper explains the drop by RtR_t becoming "less precise" (L393–394), although RtR_t is unchanged. The paper also does not state whether the ablated models were retrained or whether inputs were simply deleted at test time. If the latter, the drop only shows that the model fails on inputs it never saw during training, not that the component is useful.
  • At deployment the reasoning text goes stale; in training it never does. In training, every frame comes with reasoning written for that frame. At deployment, the reasoning model runs only when the risk level changes, and the old text is reused otherwise (L303–305), while the predicted image is updated at every step. For example, in an episode that stays SAFE, the subtask written at the start (e.g., "approach the bowl") is kept through grasping, transport, and placement. The paper's own Fig. 9 shows that delayed or misaligned reasoning lowers performance. Moreover, the aligned condition in Fig. 9 gives a CSR of 78% and 76% on Goal II/0 and II/2, which is clearly below Table 4 for the same tasks (83.2% and 82.0%), so the two may use different settings.
  • The predicted images used for training may be more accurate than at test time. The policy is trained on the world model's predictions (Table 6), and §4.1 uses one dataset for all three modules. If the predictions were generated on trajectories the world model was trained on, the model is "predicting" data it has already seen. Such predictions are far more accurate than at test time, so the policy may learn to over-rely on the image. Moreover, the world model has only seen π0.5 trajectories. Once ASR-VLA leaves those trajectories to avoid an obstacle, it reaches states absent from the training data, and the paper does not evaluate whether the predictions remain reliable there.

Minor. The hardware experiment (§4.5) is qualitative only: it reports no number of trials, no SR/CR, and no baseline.

Questions for authors
  1. Were OpenVLA-OFT and VLSA re-evaluated under the stated protocol, or were their numbers taken from the VLSA paper? How were the π0.5 numbers obtained and aggregated?
  2. What are the policy's action labels: nominal-policy rollouts, separate demonstrations, or both? Are collision-prone chunks filtered or reweighted? If the collector is the π0.5 of Table 1 (51% SR), how does unfiltered behavior cloning reach over 91% SR? Please give trajectories per scenario, the fraction with collisions, the SAFE/WARNING/DANGER proportions, and whether training configurations are disjoint from the 50 evaluation configurations.
  3. Were the ablated models retrained, or were inputs deleted at test time? In "w/o Risky-Future Image", does the reasoning model still see the predicted image? During policy training, is the reasoning regenerated for every frame, and which schedule is used in Table 1 and in Fig. 9?
  4. The same backbone fine-tuned on the same SafeLIBERO data with the same initialization and budget, but without the predicted image and reasoning (e.g., the Task-only and Task+Subtask checkpoints), and VLSA applied on top of this policy, all evaluated under one protocol on the full benchmark.
  5. Under identical observations, does the policy act differently when given opposite directional guidance (e.g., more clearance on the left vs. on the right)?
Reviewer_RMpeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ASR-VLA augments a π0.5-based manipulation policy with a future-image/risk predictor and a structured safety reasoner. The predictor is trained on nominal-policy trajectories to estimate the highest-risk future frame within a horizon, using robot–obstacle clearance to construct targets. A LoRA-adapted Qwen3-VL-8B model produces task, subtask, safety-status, and guidance fields, which condition the action policy together with the predicted image. Reasoning is refreshed when the predicted safety status changes. On 25 SafeLIBERO scenarios, the method reports 93.68% task success and 9.46% collision rate. The paper includes component ablations, reasoning interventions, perturbation tests, prediction-quality and latency analyses, and qualitative UR5e hardware demonstrations.

Critical strengths

1,A concrete integration of prospective risk evidence and action-conditioning guidance. The method supplies safety information before action generation, rather than only correcting the final action. Table 1 reports improvements in both success and the benchmark collision proxy, and the component removals indicate that the two conditioning channels provide complementary benefits. This addresses a relevant practical problem: preserving task progress while reducing unintended obstacle interaction.

2,The evaluation goes beyond a single aggregate comparison. In addition to module ablations, the paper tests fixed reasoning prompts, shuffled and delayed reasoning with matched update counts, execution perturbations, future-frame prediction quality, and runtime overhead (Section 4.4; Appendix B). The temporal-alignment control is especially useful because it tests whether relevant context matters beyond merely adding language tokens. The reported 3.51× amortized latency also makes the computational cost visible.

Critical weaknesses

1,The large improvement over baselines is not cleanly attributable to future prediction and adaptive safety reasoning. Appendix D initializes ASR-VLA from a Task+Subtask checkpoint and adds supervised adaptation, while Section 4.1 describes standard baselines as trained on LIBERO without safety-specific supervision. The phrase “same task demonstrations ... whenever their formulations permit” does not establish equal access to obstacle-augmented expert actions, phase labels, initialization, and training budget. Both single-component ablations already obtain roughly 91.6–91.7% success, much closer to the full model's 93.68% than to π0.5's 51.07%. A matched policy trained on the same safety demonstrations and subtask labels, but without either proposed channel, is essential to identify what produces the headline improvement. The existing removal ablations do not supply that control.

2,The evidence supports a narrower safety claim than action-conditioned hazard prediction. Equations 3–10 predict a future conditioned on current observations and instruction, without a candidate action, using targets from a nominal collection policy. The deployed safety-conditioned policy follows a different trajectory distribution, so the predicted frame should not automatically be interpreted as the future of its proposed action. Appendix B.5 evaluates in-domain prediction, but does not resolve this policy-shift issue. Separately, Appendix A.4 defines “collision” through obstacle displacement rather than direct contact, which can miss contacts that do not appreciably move the active obstacle. Thus, the results establish improvement on a particular proxy; they do not yet establish calibrated anticipation of the deployed policy's hazards. These limitations matter for the central safety interpretation.

Questions for authors

1,Evidence affecting attribution: What are the success, collision, and collision-free success rates of the Task+Subtask checkpoint after an equal additional training budget on exactly the same expert action data, with neither predicted future imagery nor reasoning? Please specify which demonstrations and labels each baseline receives.

2,Clarification affecting prediction validity: Were train/validation splits made by complete trajectory and initialization, rather than neighboring frames? How accurate is the risk predictor on held-out rollouts generated by ASR-VLA itself, and is its predicted future intended as a nominal-policy counterfactual or the deployed policy's actual future?

3,Clarification affecting safety measurement: How often does the displacement proxy disagree with direct robot–obstacle contacts? Please report collision-free success for the principal baselines as well as ASR-VLA, and explain whether carried-object contacts and contacts with other obstacles are counted.

4,Clarification affecting deployment evidence: How are FoundationPose-derived structured observations converted into the RGB inputs required by the world model and VLA? Does simulation pause during inference, and what is the worst-case latency of a triggered reasoning call? The 576 ms amortized figure does not determine the response delay at a hazardous transition.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.