acceptodds
Withdrawn from ICLR 2027

Probing Adversarial Vulnerabilities in Dynamic Dexterous Handover

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Dynamic Dexterous Human-to-Robot Handover (DexH2R) is an important aspect of human–robot interaction. Despite behavior-cloning methods enabling basic trajectory generation and grasping, the safety risks arising from close interaction between humans and robotic hands remain largely unexplored. In this paper, we undertake a comprehensive examination of DexH2R safety concerns by introducing different adversarial scenarios that primarily rely on 3D point clouds: 1) for grasp pose, we design a stealthy trigger to the point cloud of any object, causing the hand joints to collapse and remain closed; 2) for motion generation, we introduce two untargeted online attacks and one targeted offline attack that perturb the spatial position of the dexterous hand, preventing it from approaching the point cloud object. We conduct attacks against pre-trained victim models and introduce evaluation metrics specifically designed to assess attack effectiveness on DexH2R. Extensive experiments show that our attacks substantially reduce task success rates, with reductions of up to 100% across a range of simulated robotic tasks. Furthermore, evaluations across simulation platforms and against defense mechanisms demonstrate the strong transferability and resilience of our designed attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Adversarial Vulnerabilities, Human-to-Robot Handover

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

This paper does a relatively comprehensive examination of DexH2R safety concerns by introducing several adversarial scenarios. The authors conduct attacks against pre-trained policies. And experimental results show that their attacks seem to effectively reduce task success rates. My concerns are mainly about the physical grounding of propoesed threads and the generalization of the identified vulnerabilities.

Critical strengths
  • This paper provides the relatively comprehensive empirical evaluations across multiple failure modes, where both grasp-pose and motion-generation vulnerabilities are studied. The authors also introduces task-specific attack metrics, and several ablations and evaluations, especially with a safety-oriented human-hand intersection test. All these results make the empirical study relatively broad and systematic. But I still have concerns on the nature and the generality of the underlying threat model as discussed in critical weaknesses.
Critical weaknesses
  • My primary concern is that the physical grounding of the proposed threat is unclear. Several attacks are optimized directly in the processed point-cloud space under white-box access, making it unclear whether the resulting perturbations correspond to realistic changes in the physical world or sensing pipeline. Under such a strong intervention space, high attack success rates may partly reflect the power of the attacker, but it cannot reveal the intrinsic robustness limitation of dexterous handover policies.

  • Another concern is that the identified vulnerabilities are more likely architecture-specific, while may not be policy-general. For example, CPA relies on PointNet critical points, while the grasp attack directly optimizes the cVAE output toward an artificially chosen fist configuration. Without stronger cross-policy transfer or robustness characterization, it is difficult to tell whether these attacks reveal general failure modes of dexterous handover or mainly vulnerabilities of the specific DexH2R pipeline.

Questions for authors
  1. Can the authors provide empirical evidence that the TCA/CPA perturbations fall within errors produced by a real 3D sensing pipeline? In particular, what sensing mechanism would selectively perturb the critical points used by CPA?
  2. For the Handover-Sim experiment, is the P3TA trigger learned on DexH2R applied directly, or re-optimized for the new policy? Can the authors report quantitative zero-shot transfer results?
  3. If the attacker can directly manipulate the processed point cloud with white-box access, what additional vulnerability is revealed beyond the general fragility of the victim model?
  4. In Table 3, are the attacks adaptively re-optimized against each defense, or are previously generated adversarial examples reused?
Reviewer_KckMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates adversarial vulnerabilities in dynamic dexterous human-to-robot handover (DexH2R). The authors introduce four attacks targeting grasp pose prediction and motion generation through point-cloud perturbations. Experiments show that these attacks can significantly disrupt grasping and motion trajectories, with some achieving nearly 100% attack success rates in simulation.

Critical strengths
  1. The paper addresses an important safety problem in human-robot interaction, where adversarial failures could have serious physical consequences.

  2. The attacks cover different stages of the handover pipeline, including grasp preparation and motion generation, providing a useful evaluation of system vulnerabilities.

  3. The experimental results show strong attack effectiveness. The ablation studies and defense evaluations also provide additional insights into the proposed attacks.

Critical weaknesses
  1. The methodological novelty is somewhat limited. Most attacks build on existing point-cloud perturbation and gradient-based optimization techniques. Comparisons with stronger existing adversarial attack methods are needed.

  2. The evaluation relies heavily on simulation. Although the paper includes qualitative cross-platform results and a real-world object retrieval example, it lacks convincing physical validation in actual dexterous human-to-robot handover scenarios.

  3. Figure 5 explicitly states that its values are illustrative rather than measured results. However, Section 5.4 discusses these results as evidence of attack effectiveness under different perturbation budgets. This raises concerns about the reliability of the experimental analysis.

  4. The practical feasibility of the attacks is insufficiently discussed. In particular, the white-box assumptions, online optimization overhead, and physical realizability of point-cloud modifications deserve further investigation.

Questions for authors

Please refer to Weaknesses.

Reviewer_X8bGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates adversarial vulnerabilities in point-cloud-based dynamic dexterous human-to-robot handover. It studies two stages of the DexH2R pipeline: grasp pose preparation and approaching motion generation. The authors propose a targeted grasp attack that induces a fist configuration, together with three motion attacks based on temporally consistent transformations, critical-point perturbations, and a fixed 3D trigger. Experiments primarily evaluate these attacks in simulation, with additional defense evaluations, a qualitative transfer experiment, and a real-world object-retrieval example.

Critical strengths
  1. The paper studies an important and relatively underexplored problem: adversarial vulnerabilities in point-cloud-driven dynamic dexterous handover. Examining both grasp preparation and approaching motion provides a useful view of how corrupted observations can affect different stages of the task.

  2. The figures communicate the attack designs and resulting behaviors effectively. The quantitative evaluation is complemented by grasp and trajectory visualizations, component ablations, and defense tests, making the observed vulnerabilities easier to understand.

Critical weaknesses
  1. The assumed attacker capabilities lack a concrete deployment justification. The attacks require white-box access and direct manipulation of segmented object point clouds, but the paper does not establish how an attacker obtains this control or how the perturbation constraints correspond to practical capabilities. In particular, inserting optimized points for P3TA does not establish that a physical object can produce those observations after depth sensing, segmentation, and preprocessing. The results therefore support digital-input vulnerability more clearly than physical attack feasibility.

  2. The real-world example does not adequately validate the proposed attacks. Figure 7 concerns mobile object retrieval rather than dynamic dexterous handover, and the paper does not specify the attack used, its injection mechanism, the deployed policy, or repeated-trial results. Counting either navigation or retrieval failure as success also leaves unclear whether the failure originates in the grasp or motion-generation components studied here. Consequently, the connection between this example and the proposed methods remains insufficiently established.

  3. The real-time feasibility of the online attacks is not evaluated. TCA and CPA perform ten gradient optimization steps at each policy query, but attack latency, policy frequency, and whether computation delays are included in the closed-loop evaluation are not reported. In dynamic handover, the observations may become outdated during optimization. End-to-end timing and attack performance under the corresponding delays are needed to assess these attacks in the intended setting.

Questions for authors
  1. What concrete deployment scenario gives the attacker control over segmented object point clouds? For P3TA, can you demonstrate that a realizable physical trigger remains effective after actual depth sensing, segmentation, and preprocessing?

  2. Which attack and policy are used in Figure 7, and how is the perturbation introduced? Please clarify the failure mechanism and provide repeated-trial results that connect this example to the proposed method.

  3. What are the per-query latencies of TCA and CPA relative to the policy update interval? Does the evaluation account for object and robot motion during attack computation, and how does performance change when these delays are included?

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

Related papers

Discussion (0)

Sign in to comment.