PaperlayerView exact revision on arXiv

Source-linked AI summary

EigenDEXplore: Structured Exploration for Dexterous Manipulation with Human Priors

Harsh Gupta, Tyler Ga Wei Lum, Changhao Wang, Chuer Pan, C. Karen Liu, Jeannette Bohg, Shuran Song

arXiv:2610.07681v1cs.ROcs.AI

TL;DR

Dexterous manipulation requires coordinated hand motions that are difficult to discover through independent joint-space exploration, while learned low-dimensional action spaces can restrict expressivity. EigenDEXplore adds human-derived eigenvector perturbations to joint-space noise without changing the action representation. It consistently outperforms joint-space and learned action-space baselines across tasks and deployment settings, with larger gains under less reward shaping.

  • Problem

    Independent joint perturbations make coordinated dexterous behaviors difficult to discover, while low-dimensional human-motion action spaces can restrict the independent joint control needed for general manipulation.

    Source: §1 · pp. 1–2

  • Method

    EigenDEXplore adds perturbations along human-derived eigenvectors to independent joint-space noise while retaining joint-space actions.

    Source: §1 · p. 2, §5 · p. 9

  • Results

    Across RL, trajectory optimization, and hardware deployment, EigenDEXplore consistently outperforms joint-space and learned action-space baselines; on hardware, average task progress is 82% versus 33% for Joint-Space.

    Source: §1 · p. 2, §5 · p. 9, §4.5.1 · p. 9

  • Takeaways & Limitations

    The results suggest human-motion priors help mainly by structuring exploration rather than changing the action space, with larger gains under less reward shaping and on higher-DoF hands.

    Source: §1 · p. 2, §5 · p. 9

    The fixed linear basis cannot encode coordination patterns absent from the retargeted human data or nonlinear synergies, which must be explored through independent joint noise.

    Source: §5 · p. 9

Abstract

from arXiv · show

Dexterous manipulation poses a challenging high-dimensional optimization problem, as useful behaviors require coordinated motion across many hand joints. In reinforcement learning (RL) and sampling-based trajectory optimization, exploration commonly relies on independent robot joint perturbations, making coordinated behaviors difficult to discover. Prior work reduces this search space for grasp learning using low-dimensional spaces of coordinated joint motions learned from human hand data, but this restricts the expressivity required for general manipulation. Some combine learned and joint-space actions to restore expressivity, but this increases dimensionality and introduces redundancy. We study these effects across diverse manipulation settings, varying action dimensionality, exploration strategy, and the source of human data. Our experiments suggest that human-motion priors are most effective when used to structure exploration rather than change the action representation. Motivated by this finding, we propose EigenDEXplore, which induces correlated exploration by adding perturbations along human-derived eigenvectors to independent joint-space noise, leaving the action space unchanged. Across multiple dexterous hands, EigenDEXplore consistently outperforms joint-space and learned action-space baselines in grasping, in-hand reorientation, and contact-rich manipulation. These gains span unstructured and reference-guided RL, trajectory optimization, and sim-to-real deployment, and are largest in settings with less reward shaping and curriculum design.

1 INTRODUCTION

Dexterous manipulation requires coordinated motions that are difficult to discover with independent joint perturbations in a high-dimensional action space. EigenDEXplore uses human-derived motion patterns to structure exploration while retaining joint-space actions, and the paper reports gains across manipulation settings.

Motivation

  • Independent joint perturbations make coordinated behaviors difficult to discover, often requiring substantial computation and carefully engineered rewards, costs, or curricula.Useful contact-rich behaviors occupy only a small fraction of the hand’s high-dimensional action space.Source: §1 · pp. 1–2
  • Low-dimensional human-motion action spaces encode coordination but can restrict independent joint control, while combining them with joint actions increases dimensionality and redundancy.Prior approaches include PCA eigen-spaces and VAE latent spaces; it remains unclear whether benefits come from coordinated commands, noise-induced exploration, or changes in action-space properties.Source: §1 · p. 2

Contribution

  • EigenDEXplore adds perturbations along human-derived PCA directions to independent joint-space noise, preserving the original action representation.The comparisons suggest correlated exploration is the primary benefit, while action constraints or redundancy can hinder performance.Source: §1 · p. 2

Results

  • Across RL, trajectory optimization, and hardware deployment, EigenDEXplore consistently outperforms joint-space and learned action-space baselines across six hands.Reported gains include 34% more consecutive successes early in SimToolReal training, 40% more in DeXtreme, and a 16% reduction in SPIDER retargeting cost on average.Source: §1 · p. 2

Contribution

  • The paper also releases benchmarking code, a plug-and-play implementation, and pretrained PCA models for six dexterous hands.Source: §1 · p. 2

2 RELATED WORK

Prior work uses human hand data to guide dexterous manipulation through task guidance and coordinated action representations. This paper situates its contribution at the intersection of those representations and structured exploration.

Human priors for manipulation

  • Human-derived demonstrations and pose priors support policy supervision, initialization, and reward design, while hand and object trajectories also provide objectives for RL and trajectory optimization.These trajectories can adapt robot motions to embodiment differences and contact dynamics; human-derived goals have also been used by procedurally trained controllers at deployment.Source: §2 · pp. 2–3
  • Figure 2 illustrates the first three PCA directions after retargeting EgoSuite human hand motion to Sharpa, Wuji, Allegro, and XHand.Source: Figure 2 · p. 3

Action representations for dexterous manipulation

  • PCA-based eigen-space control builds on coordinated grasp patterns, but low-dimensional control can restrict dexterity and fine contact adjustments.Adding eigen-space residuals to joint-space actions preserves expressivity but introduces redundant action coordinates [Gupta et al., 2026b].Source: §2 · p. 3

Structure in exploration

  • Exploration interfaces can turn independent action noise into correlated joint perturbations, so human-motion priors may structure sampling as well as action representations.The study separates these effects and examines how source data and prior dimensionality influence manipulation performance.Source: §2 · p. 3

3 METHOD

EigenDEXplore retargets human hand motion to each robot, extracts coordinated joint directions, and adds noise along those directions without changing joint-space actions. The method applies this exploration strategy in both policy learning and trajectory optimization.

3.1 HUMAN-DERIVED COORDINATION PRIOR

  • The coordination prior is built by retargeting human postures to each robot, applying PCA, and retaining the first k components needed to reach a chosen variance threshold.At 90% variance, k is 9 for Sharpa’s 22 joints and 6 for XHand’s 12 joints.Source: §3.1 · pp. 3–4

3.2 ACTION SPACES AND EXPLORATION NOISE

  • EigenDEXplore adds projected eigen-direction perturbations to independent joint-space noise while keeping action dimensionality at d, unlike Eigen-Residual’s d + k action coordinates.The method rescales the two noise components to split the initial total noise variance equally with Joint-Space.Source: §3.2 · p. 4

3.3 EIGENDEXPLORE FOR REINFORCEMENT LEARNING

  • In RL, the policy predicts a joint-space command, PPO learns the policy and noise scales under the resulting sampling distribution, and deployment executes the command deterministically.The fixed eigenbasis and exploration parameters are discarded at deployment.Source: §3.3 · p. 4

3.4 EIGENDEXPLORE FOR TRAJECTORY OPTIMIZATION

  • The method samples candidate joint-target sequences with EigenDEXplore noise, evaluates them in simulation, and updates targets from the lowest-cost 10% of trajectories.Selected perturbations receive weights summing to one, with larger weights for lower-cost candidates.Source: §3.4 · p. 4
  • Figure 3 reports the highest mean consecutive successes and reward for EigenDEXplore in SimToolReal tool manipulation.It presents fixed-budget performance at 9B frames and extended training with a success-tolerance curriculum.Source: Figure 3 · p. 5

4 EXPERIMENTS

Across simulated manipulation benchmarks, EigenDEXplore improves on Joint-Space, with stronger gains when task guidance is limited and on more demanding dexterous tasks. The experiments also compare action formulations and assess how training assistance and hand embodiment affect performance.

4.1 ACTION SPACES AND EXPLORATION

  • The experiments compare four action formulations across SimToolReal and DeXtreme, reporting consecutive successes over five seeds under each benchmark’s protocol.The DeXtreme setup trains an Allegro hand for 2 billion frames with automatic domain randomization.Source: §4.1 · p. 5, §4.1.2 · p. 5

4.1.1 SIMTOOLREAL

  • At 9B frames, EigenDEXplore reaches 10.6 consecutive successes versus 7.9 for Joint-Space in SimToolReal; at 60B frames, it reaches 14.1 versus 4.9.The extended run uses a tightening curriculum, and Eigen-Space’s early advantage fades while Eigen-Residual and EigenDEXplore sustain stronger learning.Source: §4.1.1 · p. 5

4.1.2 DEXTREME

  • 12.3 vs. 8.8 mean consecutive successes: EigenDEXplore leads Joint-Space under shared randomization and progresses earlier through DeXtreme’s domain-randomization curriculum.Eigen-Space performs poorly on this task, which demands configurations that may be rare in human postures and is confined to 10 of Allegro’s 16 joints.Source: Figure 4 · p. 5, §4.1.2 · pp. 5–6

4.2 DEXMACHINA

  • With task reward alone, EigenDEXplore substantially improves object tracking on higher-DoF Allegro and XHand, while gains are limited on Inspire.With curriculum and auxiliary rewards, performance is similar to Joint-Space in most settings; Schunk on the box task improves from 0.59 to 0.84 ADD-AUC.Source: Figure 5 · p. 6, §4.2 · pp. 6–7

4.3 SPIDER

EigenDEXplore reduces SPIDER retargeting cost across hand embodiments, with larger reductions on higher-DoF hands. In real-world evaluations, it achieves higher task progress than Joint-Space on the SimToolReal tasks and more consecutive successes on DeXtreme.

4.3 SPIDER

  • 31% on Wuji and 25% on Sharpa: EigenDEXplore reduces mean retargeting cost most on these high-DoF hands, while costs on Schunk and Inspire are nearly unchanged.The reductions are relative to Joint-Space, and the fixed noise scales mean the gains come from the sampling distribution alone.Source: §4.3 · p. 7

4.5.1 SIMTOOLREAL

  • 82% vs. 33% mean task progress: EigenDEXplore exceeds Joint-Space across all four real-world SimToolReal tool-use tasks.Task progress is the fraction of demonstrated waypoints reached; the evaluation uses ten trials per object-trajectory combination.Source: Figure 8 · p. 8, §4.5.1 · p. 9

4.4 DATASET AND EXPLORATION ANALYSIS

  • At 9B frames, the EgoSuite prior reaches 10.6 mean consecutive successes, versus 9.8 for ARCTIC, 9.4 for GRAB, and 7.9 for Joint-Space.All three human-data priors improve over Joint-Space in the SimToolReal dataset comparison.Source: §4.4 · p. 8
  • At ρ = 0.90, eigen noise plus joint-space noise reaches 0.066 retargeting cost, 25% below Joint-Space; eigen noise alone at ρ = 0.50 is worse (0.126 vs. 0.089).Adding joint-space noise lowers cost at every tested threshold below ρ = 1.00.Source: §4.4 · p. 8

4.5.2 DEXTREME

  • Over 40 real-world DeXtreme trials per method, EigenDEXplore raises mean consecutive successes from 8.45 to 11.90 and the longest run from 32 to 41.Each method uses one policy trained with the 2-billion-frame budget from the DeXtreme simulation comparison.Source: §4.5.2 · p. 9

5 CONCLUSION

EigenDEXplore structures joint-space exploration with human-derived coordination patterns and outperforms joint-space and learned action-space baselines across settings. Its limitations include a fixed, retargeted linear basis and smaller improvements for low-DoF hands.

  • On real-world DeXtreme, EigenDEXplore raises mean consecutive successes from 8.45 to 11.90 across 40 trials per method.Source: Figure 9 · p. 9
  • EigenDEXplore adds human-derived eigenvector perturbations to joint-space noise and outperforms joint-space and learned action-space baselines across settings.Source: §5 · p. 9
  • Its fixed linear basis may miss absent or nonlinear synergies, depend on retargeting fidelity, and yield smaller gains for low-DoF hands.Source: §5 · p. 9

AI USE STATEMENT

The authors used generative AI tools to assist with implementation, writing, figure generation, and related-work surveying, while retaining responsibility for the paper.

  • Generative AI assisted implementation, writing, programmatic figure generation, and related-work surveying; the authors verified the code and experiments and take responsibility for the paper.Source: AI USE STATEMENT · p. 10

A ADDITIONAL ACTION SPACES

The appendix evaluates delta variants that integrate actions as increments onto previous targets. They perform comparably to absolute variants in SimToolReal but fall behind in DeXtreme.

A ADDITIONAL ACTION SPACES

B.1 SOURCE DATASET AND RETAINED VARIANCE

Across EgoSuite, GRAB, and ARCTIC priors, adding joint-space noise lowers trajectory-optimization cost across thresholds; the lowest cost occurs at 80%–95% retained variance.

B.1 SOURCE DATASET AND RETAINED VARIANCE

  • Across EgoSuite, GRAB, and ARCTIC priors, adding joint-space noise lowers cost across thresholds, with the lowest cost at 80%–95% variance retention.Source: §B.1 · p. 14

B.2 HUMAN-DERIVED VERSUS RANDOM DIRECTIONS

The section tests whether correlated exploration benefits from human-derived coordination patterns or from correlation alone. Human-derived priors match or improve on Joint-Space, while random directions perform worse; the combined eigen and joint-noise benefit follows the same trend across datasets.

B.2 HUMAN-DERIVED VERSUS RANDOM DIRECTIONS

  • Human-derived priors match or improve on Joint-Space, while random directions perform worse with eigen noise alone and with combined joint-space noise.The results indicate that correlation alone does not explain the benefit; the human-motion coordination patterns matter.Source: §B.2 · p. 14, Figure 12 · p. 15
  • Across GRAB, ARCTIC, and EgoSuite priors, the exploration-design trends remain the same, indicating that the benefit does not depend on the source dataset.Adding joint-space noise lowers retargeting cost at nearly every threshold.Source: Figure 11 · p. 15

Related papers

Loading 2610.07681v1…