- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces FAILBANK, a framework that turns runtime safety feedback into lasting policy improvements for vision-language-action (VLA) models. It addresses a key weakness of runtime shielding: while a shield can correct unsafe actions during execution, it does not update the underlying policy. As a result, the same mismatch between the policy and the shield can recur over multiple steps, reducing task completion performance. FAILBANK resolves this issue by using a fixed control barrier function (CBF)-based safety module as an observe-only teacher during rollout collection. The policy continues to act normally, while the teacher records counterfactual corrective actions whenever safety intervention would have been needed. These corrections are then filtered based on rollout outcomes, combined with successful uncorrected actions as quiet anchors, and stored in a failure bank. The accumulated data is then used to perform guarded LoRA updates on the original policy.
Critical strengths
A major strength of the paper is that it highlights an important and realistic failure mode of runtime safety systems: the persistent mismatch between a nominal policy and an external shield. The core idea of converting runtime corrections into training supervision is compelling, intuitive, and well motivated. The method is presented clearly, with a four-stage pipeline that is easy to follow and conceptually consistent. The experimental section is another strong point. It evaluates the approach across two backbones, two difficulty levels, and comparisons with both the base model and AEGIS. The paper also provides useful ablation studies that support key design decisions, particularly the choice of observe-only collection over shield-in-the-loop collection and the inclusion of quiet anchors to stabilize learning. A further practical advantage is that the adapted policy can be deployed without access to privileged geometry or a runtime shield, making the approach more appealing than methods that rely on continuous safety intervention at test time.
Critical weaknesses
Most of the evidence in the paper comes from the static-obstacle setting, so the conclusions are somewhat limited in scope. The results on dynamic obstacles are weak or unclear, which suggests that the method may not yet work well in harder safety scenarios. The approach also relies on privileged geometric information during data collection. While this makes sense in simulation, it raises questions about how well the same process would transfer to real-world robots. The comparison with AEGIS is useful in practice, but it is not a perfectly fair comparison. FAILBANK uses privileged teacher signals during training, while AEGIS depends on runtime grounding and perception. Another issue is that the method includes several hand-crafted design choices, such as admission rules, weighting bins, and guard thresholds, which may make it sensitive to implementation details. Finally, the paper focuses mostly on empirical results and would be stronger with a deeper explanation of when and why observe-only corrections lead to reliable policy improvement.
Questions for authors
First, it would be helpful to know how sensitive FAILBANK is to the quality of the CBF teacher. If the teacher is noisy, too conservative, or biased in some systematic way, does the method still improve the policy, or can it end up learning bad corrections? Second, the paper should explain more clearly how the privileged data collection setup could be replaced in a real robot setting, for example by using learned perception or rough geometry estimates. Third, since the method uses several thresholds and heuristic design choices, it would be useful to know which ones matter most and how stable the results are when they are changed. I am also curious why the main training signal uses only first-action supervision instead of full action chunks, and whether this choice is important for stability. In addition, a stronger comparison with simpler finetuning baselines would help show how much of the improvement really comes from the failure-bank design, rather than just retraining on failure cases or successful trajectories. Finally, the paper would benefit from more discussion of whether the method can be extended easily beyond flow-matching policies to token-based or diffusion-based VLA models.
Dear Reviewers,
Thank you so much for your time and constructive feedback. With the discussion deadline coming up, we’d appreciate any thoughts on our rebuttal and additional experiments. We’d be happy to answer any remaining questions and would welcome suggestions for revising the paper.
Thanks again for reviewing our work!
Best regards,
The Authors