acceptodds
Withdrawn from ICLR 2027

ReDexT: Learning a Generalizable Residual Policy for Dexterous Retargeting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Dexterous retargeting often requires fitting robot controls to each human hand-object motion, making large collections expensive to process. Offline inverse kinematics provides inexpensive nominal controls, but cannot correct errors caused by contact and object dynamics. We present ReDexT, a two-stage reinforcement-learning framework that learns one residual policy to execute new trajectories from inverse-kinematics commands. The key is to learn reusable feedback by varying both the commands the policy corrects and the motions it practices. We first perturb successful source commands while keeping their motion targets fixed, so the policy learns to correct varied execution errors. Rollout-guided training then broadens trajectory coverage through fresh on-policy learning, without treating rollout actions as imitation labels. Once trained, the shared policy executes new trajectories with frozen weights and supports optional local or shared adaptation for better tracking. On new trajectories screened for initial stability in single-hand simulation, frozen ReDexT achieves higher success and shorter processing times than the evaluated per-trajectory methods. Shared adaptation further raises success from 55.08% to 80.08% across 256 test trajectories. Ablations show that control perturbations ease transfer to kinematic commands, and expanded IK training broadens success coverage. The training recipe extends to four additional robot hands, each with a separately trained policy.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Dexterous Manipulation, Motion Retargeting, Residual Reinforcement Learning, Zero-Shot Generalization, Policy Adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_SHS2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ReDexT learns a shared residual policy through control perturbations around optimized seed commands, followed by rollout-guided training over IK priors. On 256 held-out trajectories, it achieves 55.08% frozen success and 39.06% horizon completion, increasing to 80.08% and 75.39% after shared adaptation; the training recipe is also evaluated with separately trained policies for four additional hands.

Critical strengths

Control perturbations substantially improve transfer to IK commands, with consistent held-out gains under both SPIDER and IK training priors (Tables 8–9). This provides useful evidence for learning feedback that is less dependent on its nominal control source.

The Full-pool comparison matches architecture, control scales, training data, and interaction budget, supporting an improvement from 49.22% to 55.08% frozen success. The additional-hand experiments and explicit separation of frozen execution, adaptation, and horizon completion strengthen the empirical evaluation.

Critical weaknesses

The added value of the distinctive training stages remains unclear. Noisy IK and SPIDER bootstrap both achieve 28.13% held-out success (Table 8), while skipping bootstrap achieves 39.06% versus 40.23% for the bootstrap-plus-expansion route despite fewer updates (Table 11). Since these ablations use different control scales from the final model, they do not establish whether optimized-source bootstrap and rollout-based selection improve upon a simpler IK training curriculum at the default settings.

The evidence for an advance over existing shared trackers is incomplete. Prior methods such as DexH2R and DexTrack already learn shared reference-conditioned feedback, whereas the external comparisons primarily concern per-trajectory fitting; Appendix B.3 explicitly identifies IK-RL as an internal baseline rather than a reproduction. The matched Full-pool comparison is informative, but its 5.86-percentage-point improvement is based on one training seed, leaving the reproducibility of this gain unresolved.

Questions for authors

At the default control scales, how does ReDexT compare with IK-based bootstrap followed by the same expansion procedure, curriculum resets, and total interaction budget? This would clarify the contribution of optimized source commands beyond selecting successful seed tasks.

Can the evaluation include a representative existing shared tracker under a controlled protocol? This would help establish the success–compute tradeoff relative to methods that already avoid fitting each new trajectory.

How stable is the ReDexT–Full-pool improvement across training seeds? The paired intervals in Table 9 address different comparisons and quantify trajectory variation conditional on fixed checkpoints, rather than variability across training runs.

Reviewer_6p2SOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ReDexT, a framework for learning a shared residual RL controller to improve the physical feasibility of retargeted human hand-object trajectories. The method first trains a residual policy using SPIDER-optimized trajectories with control perturbations, then expands training to additional trajectories using IK-based controls and rollout-guided task selection. The resulting controller can execute unseen reference trajectories without per-trajectory optimization, while optional local or shared RL adaptation further improves tracking performance. Experiments across five simulated robot hands demonstrate improved tracking success and reduced per-trajectory processing time compared with the evaluated baselines.

Critical strengths
  • Clear presentation and informative supplementary materials. The paper is well organized with clear explanations of the method and experiments. The supplementary website provides useful visualizations of the retargeted trajectories and qualitative comparisons.

  • Important and well-motivated problem formulation. Generating physically feasible robot trajectories from human motion data is a valuable problem, particularly for scaling robot policy learning from human demonstrations. Learning a reusable controller that amortizes expensive per-trajectory optimization is a promising direction, and the proposed shared residual RL formulation is very sensible.

Critical weaknesses
  • Questionable external baseline results and potentially flawed implementations. DexMachina and ManipTrans achieve surprisingly low success rates (15.63% and 14.06%, respectively), barely outperforming the IK-only baseline (12.50%), despite extensive training. This raises concerns about potential implementation or configuration issues. Although adapting these methods to another hand should be feasible in principle, the paper provides insufficient validation that the adaptations work correctly. The paper also did not offer much insight to explain why the results do not match with the capabilities demonstrated by the original methods

  • Unfair comparison of methods with different objectives and capabilities. ReDexT is a reference-conditioned tracking controller that requires precomputed IK trajectories. In contrast, the original DexMachina and ManipTrans frameworks support learned policies and more complex manipulation settings. ReDexT's reported processing time excludes its substantial upstream RL training, which should be considered when interpreting the efficiency comparison.

  • Limited task complexity and diversity. The evaluation consists entirely of single-handed rigid-object manipulation trajectories. These are considerably narrower than the bimanual, articulated-object, and long-horizon manipulation tasks addressed by the original baselines. Although the benchmark includes many trajectories, they largely represent variations of the same object-tracking problem rather than diverse manipulation skills. It's unclear whether the controller can handle more complex contact interactions, articulated objects, or coordinated bimanual manipulation.

  • No real-world validation. The simulation also uses position servos without explicit actuator force/torque limits or arm kinematics.

Questions for authors
  1. Baseline implementation and validation: How exactly were DexMachina and ManipTrans implemented and adapted for this benchmark? Were official implementations used? Can the authors provide training curves, convergence diagnostics, and qualitative rollouts demonstrating that these baselines function correctly? What explains their extremely low success rates despite substantial per-trajectory optimization?

  2. Generalization to complex manipulation: How would the proposed framework handle articulated-object manipulation, multi-object interactions, or long-horizon bimanual tasks? In particular, can one shared residual controller handle contact-rich tasks involving multiple objects, changing contact modes, and coordinated hand motions, or would these require substantial changes to the policy architecture and training procedure?

Reviewer_1aGMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ReDexT, a two-stage reinforcement-learning framework for dexterous motion retargeting. Rather than optimizing a separate controller for every new hand–object trajectory, it trains a shared residual policy that corrects inexpensive offline inverse-kinematics (IK) commands using feedback from the current execution state and future motion reference. The first stage bootstraps the policy on successful SPIDER-generated control sequences, applying temporally correlated, curriculum-adjusted perturbations to the nominal controls while keeping the reference motion fixed. The second stage expands training to additional trajectories: clean-IK rollouts identify successful candidates and selected failures, and PPO continues on fresh on-policy experience around the original IK controls, without using rollout actions as imitation targets.

On 256 held-out, initial-stability-screened trajectories in single-hand simulation, the frozen policy achieves 55.08% success, versus 15.23% for IK-only and 49.22% for matched-budget full-pool training. Joint adaptation on the target trajectories increases success to 80.08%. The paper also compares against per-trajectory methods on a 32-trajectory subset and demonstrates the training recipe on four additional simulated hands, each with a separately trained policy.

Critical strengths

1. A practically useful shared-controller formulation with a meaningful controlled comparison. ReDexT can execute held-out trajectories using only offline IK and a frozen feedback policy, avoiding target-specific policy optimization at deployment. On the full 256-trajectory set, it improves success from 15.23% (IK-only) to 55.08%, and exceeds matched-budget full-pool IK training by 5.86 percentage points (Table 1). These results suggest that the overall training recipe can improve the coverage of a reusable dexterous tracker, rather than merely improving accuracy on already-successful motions.

2. Useful diagnostic experiments and multi-hand applicability. The source-to-IK control-switch experiment (Table 2A) shows that nominal-control perturbation substantially reduces the success-rate drop when replacing SPIDER controls with IK, from 30.73 to 3.65 percentage points. The appendix reports further controlled variants, failure cases, and separate results on five simulated hand embodiments. This provides valuable practical evidence about which aspects of the training setup matter.

Critical weaknesses

1. The incremental value of the distinctive two-stage recipe is not convincingly isolated. The strongest bootstrap ablation shows that SPIDER + noise and IK + noise obtain exactly the same 28.13% held-out success after 50k updates (Appendix Table 8), leaving the need for successful SPIDER source controls unclear. Likewise, the proposed lowest-error failure selection achieves 40.23% frozen success, versus 41.02% for random-failure selection and 39.45% when using all candidates (Appendix Table 11). These small single-run differences do not establish that the proposed rollout-guided selector is preferable to simpler alternatives. Although the complete recipe beats full-pool training (55.08% vs. 49.22%), that comparison changes several factors simultaneously and does not identify which component produces the improvement. This matters because the claimed methodological contribution is primarily the training strategy, not a new RL algorithm or policy architecture.

2. The scope and cost of the reported gains are insufficiently established for the broader efficiency/generalization claims. The principal matched-budget full-pool comparison is reported for one training seed; uncertainty across independent RL training runs is not characterized. The external per-trajectory comparisons use only 32 trajectories, and some baselines retain different embodiments, simulators, or evaluation horizons; there is no direct, controlled reproduction of a strong existing shared residual tracker such as DexTrack. The reported approximately 0.8 minutes per trajectory measures post-IK processing after training, excluding approximately 174.13 hours of upstream shared-policy training and the cost of producing the SPIDER source controls (Appendix B.5). Furthermore, frozen success is 75.32% on seen-object versus 24.51% on unseen-object trajectories, and all results are in simulation with initial-stability-screened tasks. These qualifications narrow what can be concluded about end-to-end scalability and generalization.

Questions for authors

Q1 — Necessity of SPIDER-source bootstrap (potentially assessment-changing). Could the authors provide a fully matched IK-bootstrap → IK-expansion experiment against SPIDER-bootstrap → IK-expansion, holding the seed/candidate trajectories, total interactions, noise schedule, action ranges, and selection procedure fixed, ideally over multiple training seeds? Appendix Table 8 reports identical 28.13% held-out success for SPIDER+noise and IK+noise at the bootstrap endpoint, so the added value of the costly source controls is not yet evident.

Q2 — Contribution of the proposed curriculum and candidate selector (potentially assessment-changing). Could the authors separate the effect of the adaptive noise curriculum from that of fixed-amplitude temporally correlated nominal-control noise, and evaluate the lowest-error candidate selector against random-failure selection over multiple training seeds and matched task-exposure budgets? In the reported ablation, random-failure selection is slightly better (41.02% vs. 40.23%), so it is unclear whether the specific selection mechanism contributes beyond broader IK task exposure.

Q3 — Interpretation of efficiency and baseline comparisons. Can the authors report an estimated end-to-end break-even point that includes source-control generation, upstream training, and offline IK, rather than only post-IK test-time processing? It would also help to clarify which conclusions remain valid under directly matched shared-policy baselines and to report variance across independently trained policies. These details would materially affect the strength of the efficiency claim and the comparison with prior generalist tracking approaches.

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

Related papers

Discussion (0)

Sign in to comment.