acceptodds
Withdrawn from ICLR 2027

DexPolicyEvo: Evolution of Dexterous Grasping Policies through Failure-Driven Self-Improvement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Enabling dexterous grasping policies to continually improve toward open-vocabulary object grasping remains an open challenge. Achieving this efficiently requires understanding a policy's current capabilities and weaknesses to tailor its training accordingly. With this goal in mind, we introduce DexPolicyEvo, a system that enables dexterous grasping policies to evolve autonomously through failure-driven self-improvement. Its key idea is to connect failure diagnosis with targeted training-data construction and geometric curriculum design through measurable and editable object properties. Six collaborating agents implement this improvement loop: the Diagnostic Designer and Policy Evaluator construct tests across everyday scenarios and assess policy capabilities; the Failure Analyst and Repair Designer identify weaknesses and devise targeted training data and curricula; and the Policy Trainer and Policy Verifier update the policy and validate its improvement. As the policy improves, the agents revise its training plan to address emerging weaknesses and further expand its grasping capabilities. Experiments demonstrate that DexPolicyEvo continually improves policy generalization while preserving existing grasping capabilities.

Primary area:
applications to robotics, autonomy, planning
Keywords:
DexPolicyEvo: Evolution of Dexterous Grasping Policies through Failure-Driven Self-Improvement

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

This paper introduces DexPolicyEvo, a multi-agent framework for failure-driven self-improvement of dexterous grasping policies. It identifies measurable geometric failure factors, generates targeted training objects, and introduces Bridge curriculum for difficult grasps. Experiments on UniGraspTransformer and ResDex demonstrate substantial improvements while largely preserving existing capabilities.

Critical strengths
  • The paper addresses an important limitation of conventional large-scale dextrous grasping: indiscriminately expanding training datasets may be inefficient because the resulting data do not necessarily target the weaknesses of the current policy.
Critical weaknesses
  • The reproducibility statement mentions supplementary materials, but none are provided.
  • The roles of the six agents are not evaluated in the ablation study, making the design seem overly complex and potentially costly
Questions for authors
  • Are there any experiments with vision-conditioned policies? Both UniGraspTransformer and ResDex have been evaluated in vision-conditioned settings.
  • How scalable is the method? The experiments only report three rounds of evolution. Would more compute and time lead to consistent improvements?
  • How stable is the autonomous evolution process? Are the results reproducible across different agent generations, diagnostic sets, and training seeds? How often do repair attempts fail or candidate policies get rejected?
  • The Bridge curriculum uses predefined training budgets, success thresholds, and a maximum number of intermediate levels. How sensitive is the performance to these design choices? Would fewer Bridge levels, different geometric step sizes, or alternative progression thresholds achieve similar improvements?
Reviewer_ZVYgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces an agent orchestration framework for self-improvement of dexterous grasping policies. It decomposes the process into six agents that diagnose, analyze, repair, and update the policy iteratively. The diagnostic agents evaluate the policy on different objects and pass the results to an analyzer agent, which identifies object properties associated with failures. A repair agent uses this analysis to construct targeted training objects and a curriculum. The update agents then train the policy and verify its improvement and retention of existing capabilities. The framework is evaluated on UniGraspTransformer and ResDex in simulation, with results showing improved grasping performance over successive iterations.

Critical strengths

None

Critical weaknesses

The proposed agent organization is not empirically justified. The paper decomposes policy improvement into six fixed agent roles, but does not ablate this organization or compare it against simpler alternatives. For example, could a single agent with the same tools, feedback, and resource budget achieve comparable results when prompted to improve the policy? Does separating diagnosis, analysis, and repair into different agents improve performance over executing the same workflow with one agent? The experiments do not establish what this particular orchestration contributes, despite its central role in the proposed framework. They do not explain why the proposed organization is effective, which aspects of agent collaboration matter, or how that collaboration accounts for the observed improvements. Even the choice of underlying agent model is not ablated.

The experimental design does not establish a clear scientific takeaway about the proposed framework. The paper compares class-matched random sampling, failure-neighborhood generation, and curriculum variants, but does not sufficiently explain why these are the relevant alternatives, whether they reflect approaches used in prior literature, or why these comparisons were chosen over others. These ablations focus on training-data construction and curriculum design, despite agent orchestration being a central claimed contribution. It is unclear what broader insight these comparisons are intended to provide about the framework beyond whether particular implementation choices improve performance.

Questions for authors

Can the authors provide a systematic ablation of the agentic architecture? The evaluation should establish which components and interactions contribute to performance, whether the prescribed division of responsibilities and coordination structure are justified, and why the organization works.

How do the ablation variants relate to relevant prior work on failure-directed learning and curriculum design? Please clarify which existing approaches they represent and how those connections inform the interpretation of the results.

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

The paper introduces DexPolicyEvo, a new multi-agent framework for improving existing dexterous grasping policies in a simulation setting via failure-driven self-improvement. The main contribution is a failure-driven self-improvement framework that is guided by grasping failures and measurable and editable geometric properties of the simulated objects used to guide targeted training-data generation and geometric curriculum construction. The authors evaluate the framework using two existing policies, UniGraspTransformer and ResDex, over three improvement loops in a simulated environment using Isaac Gym and a Shadow Hand robotic hand. From this evaluation, authors observed better performances with difficult objects, preservation of performance on the original benchmarks and favorable results in the ablation studies.

Critical strengths
  • Substantial improvements on challenging objects with two different policies while preserving or improving performance on the original benchmarks. The experiments in this paper demonstrate concrete empirical evidence of the framework’s effectiveness across different policy-learning approaches.
  • The use of a failure-driven, geometry-informed heuristic is an interesting approach for targeted training-data generation. The paper illustrates well the importance of carefully designing how learning systems are improved with concrete measurable factors rather than relying on less targeted data generation strategies.
Critical weaknesses

The ablation studies demonstrate the benefits of targeted data generation and the Bridge curriculum but do not clearly isolate the contribution of the multi-agent architecture itself. A comparison with a simpler adaptive pipeline could have helped determine whether the complexity added by the multi-agent architecture is justified.

Questions for authors

Could the authors clarify what specific benefits the multi-agent architecture provides compared to a simpler adaptive pipeline using the same failure-guided data generation and curriculum learning mechanisms?

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

Related papers

Discussion (0)

Sign in to comment.