acceptodds
Under review as a conference paper at ICLR 2027

Risk-Aware Benchmarking and Adaptation for Gentle Retrieval Through Yielding Occlusions

Abstract

Robots retrieving objects from cluttered environments often must interact with compliant structures that occlude a target while avoiding damage to both the surroundings and the robot. Existing manipulation benchmarks and policies typically emphasize task completion or aggregate interaction costs, making it difficult to isolate how occluder mechanics, retrieval stage, and rare contact events govern failure under yielding occlusions. We introduce GRAM, a benchmark for gentle retrieval through yielding occlusions that factorizes retrieval into Expose, Attach, and Transport stages, procedurally generates thin-shell and branching occluders with controlled morphology shifts, and defines success under explicit environment-disturbance and robot self-preservation budgets. Across modern visuomotor and vision-language-action policies, GRAM reveals three recurring failure modes: mechanics-driven distribution shifts alter effective interaction strategies, bottlenecks move across retrieval stages, and budget violations are often caused by short peak events rather than aggregate interaction cost. Motivated by these findings, we propose SRAH, a lightweight residual adapter that conditions a frozen policy on predicted retrieval stage, recent mechanics response, remaining budget headroom, and short-horizon hazard, while using unsafe and failed rollouts as counterexample supervision. Across ACT-style and OpenVLA backbones and both occluder families, SRAH improves budgeted gentle success by 4.1–5.7 percentage points under cross-family and tightened-budget evaluation while reducing environment and robot-side violations. Targeted real-robot analogues exhibit the same qualitative pattern, improving acceptable retrieval more than raw completion and suggesting that benchmark-derived structure can guide practical adaptation for damage-sensitive, contact-rich retrieval.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Yielding occlusions, mechanics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_43nGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper focus on an interesting and challenging task. The main contributions include a risk-aware diagnostic benchmark and a benchmark-derived adaptation framework. The benchmark demonstrate that the gentle retrieval failure mainly comes from three interaction-state variables. And SRAH framework employs the interaction states on the frozen policy for residual correction, where the failed trajectory is used for instructing the policy what "danger looks like". The paper provided relatively comprehensive validation, but may contain some overclaims and presentation issue.

Critical strengths
  • This paper aim to address the gentle retrieval through yielding occlusions, which is an interesting, practical and challenging task in robot manipulation. Therefore, such benchmark and further adaption framework is timely and meaningful at current stage. The benchmark, experiments, and adaption methods are overall reasonable and relatively comprehensive.
  • The author's design of the risk metric in the benchmark was reasonable and verified through manual annotation. Even though it contained pre-defined weights, the scanning of weights also reduced my concerns about sensitivity.
Critical weaknesses
  • The benchmark itself appears novel in its specific setting, but the paper does not sufficiently position its contributions against recent deformation-aware and contact-rich manipulation benchmarks (e.g. SoftVTBench, MoDeSuite, Tabero). Adding a discussion on such works would help reduce some overclaiming and focus on the core contributions.
Questions for authors
  • Did you use a separate development set to derive the GRAM findings that motivated SRAH? The paper clearly separates train/val/test for model fitting, but it is unclear whether the SRAH design was informed by the final evaluation set
  • Does hazard gating actually prevent budget violations? Fig.4 shows that the gate decreases near risky events, but this does not mean that those events are avoided. A controlled comparison with and without hazard gating would help clarify
  • Is there a plan to expand the coverage of the benchmark? Currently, only two types of gentle retrieval scenarios are being discussed
  • Expressions are highly repetitive and overly defensive. The introduction repeatedly summarizes contributions, which makes the paper unnecessarily difficult to read
2 comments

Thanks for authors' rebuttal, especially for the additional experiments and contribution clarification.

However, considering the relatively narrow coverage of the benchmark -- only two complementary retrieval scenarios. I cannot give a higher rating than current evaluation, therefore will maintain my original rating.

Overall, I would encourage authurs enrich this research with more comprehensive evaluation in more diversified "Gentle Retrieval through Yielding Occlusions" scenarios.

Thank you for acknowledging our clarifications and experiments, as well as identifying ways we could strengthen the research. This is a lot of work and we're really grateful for it!

Reviewer_F4A3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies robotic object retrieval through deformable occlusions, where successful manipulation requires limiting environmental disturbance and excessive robot-side contact loads, beyond task completion. The authors introduce GRAM, a benchmark that evaluates retrieval using environment-disturbance and robot self-preservation budgets. GRAM includes two families of deformable occlusions, thin-shell folds and branching corridors. And it shows that retrieval failures depend on local interaction mechanics, retrieval stages, and short-duration high-risk contact events.

Motivated by these findings, the authors propose SRAH, a residual adapter for frozen visuomotor policies. The adapter is trained to imitate gentle successful demonstrations, while unsafe trajectories provide auxiliary supervision for interaction-state and risk prediction. Experiments show that SRAH improves gentle retrieval success under cross-family transfer compared with standard imitation-learning policies that do not explicitly account for disturbance costs.

Critical strengths
  • The paper addresses an important problem where many manipulation tasks require careful physical interaction beyond simply achieving task completion. This is relevant when robots operate in delicate or high-value environments or in close proximity to humans.
Critical weaknesses
  • The usefulness of the proposed benchmark is not demonstrated as convincingly as it could be. Much of its value is illustrated through SRAH, a method specifically designed to leverage the additional disturbance-related information introduced by GRAM. I believe the key value of GRAM lies not merely in introducing binary gentle-success labels, but in providing continuous, fine-grained measurements of environmental disturbance and robot-side costs.

    • An important missing baseline is standard imitation learning trained exclusively on gentle-success demonstrations. This would help determine whether the observed improvements come from filtering out unsafe demonstrations or leveraging the richer disturbance information.
  • The methodological contribution of SRAH appears limited, as it primarily combines residual imitation learning with auxiliary predictions. Its deployment requires online estimates of environmental disturbance and robot-side costs, which may not be available in practical settings. I believe SRAH would be more convincing if positioned as a means of demonstrating the learning benefits of GRAM's continuous disturbance measurements, instead of a standalone contribution.

    • Importantly, the current evaluation does not clearly specify the composition of the training demonstrations used for the baseline imitation policies. If these baselines are trained on unsafe completions, while SRAH uses only gentle-success demonstrations for residual imitation, the observed improvements would be less surprising
Questions for authors
  • Do the training datasets for the baseline policies include failed trajectories?

  • Can the authors include a baseline trained exclusively on gentle-success demonstrations?

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

This paper tackles the problem of robotic retrieval of objects through compliant occlusion. The authors argue that existing benchmarks don't provide insights into failures and, hence, propose a new benchmark, GRAM. GRAM provides diagnostics of different stages of the task: Expose-Attach-Transport. Through insights from the benchmark, the authors propose a method for adaptive policy method, SRAH, which shows 4.1-7.1% improvement.

Critical strengths
  • New benchmark
Critical weaknesses
  • Providing a task stage breakdown alone doesn't provide enough value as a benchmark
  • Improvement from diagnostics is minimal (4.1-7.1%), especially considering that the success rates are low (<60%)
  • The paper is, in general, difficult to follow. E.g., "GRAM shows that retrieval failures are structured by functional stage, local mechanics response, and short-horizon peak-event risk". This is difficult to understand.
Questions for authors

How does the new benchmark compare against existing benchmarks for similar tasks? It wasn't clear from the related work section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.