PaperlayerView exact revision on arXiv

Source-linked AI summary

ReDex: Repairing Sim-to-Real Dexterous Policies by Finger-Level Compliant Interaction

Jinzhou Li, Hadi Tabatabaee, Kelin Yu, Yuyin Sun, Cheng-Hao Kuo, Roberto Martín-Martín, Nima Fazeli, X. Alice Wu, Xianyi Cheng

arXiv:2610.07525v1cs.RO

TL;DR

Sim-trained dexterous policies can fail at real-world contact placement, timing, and force regulation, while retaining useful coordination. ReDex corrects selected fingers through compliant physical interaction and trains a force-conditioned policy from force-informed rollout targets. Across evaluated objects, it reports improved task performance with limited correction data.

  • Problem

    Dexterous correction methods leave a gap in easily adjusting individual finger motion and contact forces, while operators typically lack direct perception of robot contact forces.

    Source: §I · p. 1

  • Method

    ReDex combines selective compliant finger correction with force-informed target reconstruction and behavior cloning to train a standalone force-conditioned policy.

    Source: §I · pp. 1–2

  • Results

    Mean rotation progress improves from 33% to 99% on the small screwdriver, from 19% to 92% on the large screwdriver, and cellphone flipping success increases from 12% to 92%.

    Source: §V · p. 6

  • Takeaways & Limitations

    Across two skills and five objects, ReDex improves real-world performance with limited correction data and without tactile simulation or additional simulation training.

    Source: §VI · p. 8

Abstract

from arXiv · show

Dexterous manipulation policies trained in simulation often fail to transfer to the real world because of errors in contact timing and force regulation. Yet these policies can retain useful multi-finger coordination for task progression. We propose ReDex, a framework for adapting a simulation-trained base policy to the real world by correcting local contact failures and incorporating tactile feedback. Starting from a proprioception-only base policy, ReDex allows a human operator to physically correct contact failures at selected fingers under compliant control during real-world rollouts, while the frozen base policy continues to control the remaining fingers. These rollouts combine base policy execution, human-corrected finger motion, and fingertip force observations. We reconstruct force-informed targets from these rollouts to train a standalone force-conditioned policy via behavior cloning. This design reduces human correction effort, enables learning of contact regulation from real-world interaction, and introduces force feedback into a proprioception-only policy without tactile simulation or complex full-hand teleoperation. We evaluate ReDex on two challenging, contact-rich dexterous manipulation tasks on real hardware. Compared with sim-to-real transferred base policies, ReDex increases Object Flipping success rate from 14\% to 86\% across two objects and average Screwdriver Rotation progress from 26.0% to 95.3% across three objects.

I. Introduction

Sim-to-real dexterous policies can retain useful finger coordination while failing at local contact placement, timing, and force regulation. ReDex corrects selected fingers through compliant physical interaction and uses real-world force feedback to train a standalone policy.

I. Introduction

  • ReDex addresses sim-to-real failures in contact placement, timing, and force regulation that are difficult to resolve through further simulation retuning.Repeated reward or domain-randomization retuning can be costly without necessarily resolving observed contact failures.Source: §I · p. 1
  • Existing correction methods leave a gap in easily correcting both individual finger motion and contact forces during dexterous manipulation.Pose correction alone does not ensure suitable contact forces, and operators typically lack direct perception of the robot's contact forces.Source: §I · p. 1
  • ReDex lets an operator physically correct only failing fingers under compliance while the frozen base policy controls the remaining fingers.This preserves simulation-learned multi-finger coordination and limits the required human effort.Source: §I · pp. 1–2
  • ReDex reconstructs force-informed joint-position targets from corrected rollouts and distills them into a standalone force-conditioned policy via behavior cloning.The rollouts record base-policy behavior, human corrections, and fingertip forces; evaluation covers Object Flipping and Screwdriver Rotation.Source: §I · pp. 1–2
  • The contributions include a finger-compliance correction framework, a relabeling pipeline for dense supervision, and improved force-conditioned policies from a few minutes of correction per object.The relabeling combines corrected motion, real-world tactile feedback, and base-policy continuation.Source: §I · p. 2

II. Related Work

Simulation enables large-scale dexterous policy training, but transfer remains challenging. ReDex preserves learned multi-finger coordination while correcting selected fingers on real hardware.

A. Learning Dexterous Manipulation via RL

B. Learning from Human Correction

Prior correction approaches differ in how directly they control fingers and provide contact-force information. ReDex adapts a frozen simulation-trained policy through physical finger correction, target reconstruction, and policy learning.

B. Learning from Human Correction

  • Unlike arm-focused correction methods, ReDex intervenes only on failed fingers while the base policy controls the rest, preserving learned coordination during correction.Earlier dexterous approaches either do not allow direct individual-contact correction or require the operator to coordinate the hand.Source: B. Learning from Human Correction · p. 2

C. Learning with Contact Feedback

  • ReDex uses measured real contact both to reconstruct force-informed supervision and as policy observations, without requiring full-hand force-rich demonstrations.The related work contrasts this with approaches requiring sensor-specific tactile modeling and calibration in simulation.Source: C. Learning with Contact Feedback · p. 2
  • The pipeline trains a proprioception-only policy in simulation, collects compliant finger-level corrections while the base controls remaining fingers, then trains a standalone force-conditioned policy.Measured fingertip forces map to force-informed joint targets, which are blended with evolving base-policy targets at intervention onset and release.Source: Figure 2 · p. 3

III. Method

  • ReDex adapts a frozen simulation-trained policy through physical correction via finger compliance, action-target reconstruction, and policy learning.The objective is to correct local finger behavior while retaining base-policy control of the remaining fingers.Source: §III · p. 3

A. Sim-to-Real Policy Training

The base policy is trained in simulation and deployed using proprioception. It is trained once per task and reused across the corresponding real objects.

A. Sim-to-Real Policy Training

  • The base policy generates joint-position targets from a history of joint positions and previous targets, using proprioception at deployment.Source: A. Sim-to-Real Policy Training · p. 3
  • Training uses a privileged PPO teacher followed by an adaptation module that estimates its latent from joint-position and target history.The resulting policy requires only proprioception at deployment.Source: A. Sim-to-Real Policy Training · p. 3
  • Screwdriver Rotation uses an object angular-velocity objective, while Object Flipping uses a keypoint-based goal-pose objective.Source: A. Sim-to-Real Policy Training · p. 3
  • Task-specific domain randomization covers object dynamics, controller gains, and observation and action noise; each base policy is trained once and deployed on corresponding real objects.Source: A. Sim-to-Real Policy Training · p. 3

B. Finger-Level Compliant Correction

ReDex lets an operator physically correct selected fingers while the base policy continues controlling the remaining fingers. Selective compliance and joint-target overrides support guidance without disrupting task execution.

B. Finger-Level Compliant Correction

  • Selected joints use compliant guidance targets while unselected joints track base-policy targets, allowing physical correction without disrupting task execution.The base policy continues receiving measured states, including corrected finger positions.Source: B. Finger-Level Compliant Correction · pp. 3–4
  • The operator selects fingers through keyboard commands, which define a binary joint-level mask for compliant versus nominal control.Selected joints also use reduced torque limits and a dedicated derivative-gain profile.Source: B. Finger-Level Compliant Correction · p. 3
  • Guidance targets initialize at measured joint positions and follow the measured configuration during intervention; releasing a finger restores nominal settings and base-policy targets.Corrected positions remain in the full-hand history and can influence base-policy outputs for other fingers.Source: B. Finger-Level Compliant Correction · pp. 3–4
  • Rollouts record policy observations, measured joint positions, base-policy targets, fingertip forces, and intervention masks for reconstructing force-informed targets.Measured positions capture operator-induced correction motion.Source: B. Finger-Level Compliant Correction · p. 4

C. Reconstructing Force-informed Action Targets

ReDex combines corrected joint motion with measured fingertip forces to reconstruct force-informed targets. Force-derived offsets are local to corrected fingers, and reconstructed targets replace base-policy commands only for guided joints.

C. Reconstructing Force-informed Action Targets

  • Force-informed targets combine the corrected joint configuration with an offset derived from measured fingertip forces, since physical correction captures motion but not desired contact force.For corrected fingers, resultant forces are thresholded and expressed in the hand base frame.Source: C. Reconstructing Force-informed… · p. 4
  • Measured fingertip forces are converted to joint-position offsets through a quasi-static virtual-spring model, then added to measured joint positions.The conversion uses the fingertip translational Jacobian and a fixed diagonal virtual joint-stiffness matrix.Source: C. Reconstructing Force-informed… · p. 4
  • The force-derived offset remains local to corrected fingers, and γ = 0.05 is fixed for constructing all policy training targets.The scale was selected in preliminary replay experiments based on observed contact maintenance and motion fidelity.Source: C. Reconstructing Force-informed… · p. 4
  • The reconstructed target combines corrected motion with a force-derived offset, whose linearized virtual Cartesian displacement is defined at the current configuration.The passage identifies the offset as the motion component augmented by measured contact force.Source: C. Reconstructing Force-informed… · p. 4
  • During intervention, force-informed targets replace base-policy targets only for corrected joints; unselected joints retain their base-policy targets.The policy remains in closed loop, and switching between target sources uses minimum-jerk interpolation around intervention boundaries.Source: C. Reconstructing Force-informed… · p. 4

D. Learning Force-Conditioned Policy

ReDex distills rollouts into a standalone force-conditioned policy trained by behavior cloning. At deployment, it predicts joint-position chunks from recent joint-position and fingertip-force histories.

D. Learning Force-Conditioned Policy

  • The standalone policy maps joint-position and fingertip-force histories over H = 5 steps to full-hand joint-position chunks of length N = 16.The policy uses action-chunking behavior cloning.Source: D. Learning Force-Conditioned Policy · p. 4
  • The force-conditioned policy is trained by optimizing the behavior-cloning objective.The cited passage identifies the procedure as standard behavior cloning.Source: D. Learning Force-Conditioned Policy · p. 4
  • At deployment, the policy executes the first 12 targets of each 16-step chunk at 20 Hz before predicting the next chunk.Predictions use joint-position and fingertip-force histories.Source: D. Learning Force-Conditioned Policy · p. 4

IV. Experiments

The experiments assess real-world performance across two in-hand manipulation skills and five objects, alongside tactile supervision, transition blending, and human correction effort. They also include qualitative rollouts of autonomous execution after human corrections.

IV. Experiments

  • The evaluation spans Object Flipping and Screwdriver Rotation on five real objects, with component and representation studies on cuboid flipping.The experiments examine performance, tactile information and force-informed supervision, and the human effort required for correction.Source: §IV · p. 4
  • The experiments ask whether ReDex improves real-world performance and how tactile observations, force-informed targets, transition blending, and correction effort contribute.Source: §IV · p. 4
  • Fig. 3 shows four frames of autonomous execution per rollout, comparing object-flipping and screwdriver-rotation examples across objects.Frames run top to bottom; screwdriver insets show orientation changes, and T marks elapsed seconds.Source: Figure 3 · p. 5

A. Experimental Setup

The setup uses a tactile-sensing five-finger hand and evaluates two manipulation skills with task-specific metrics. It trains object-specific policies from real-world correction data and compares against direct base-policy execution and residual-correction baselines.

A. Experimental Setup

  • The RobotEra XHAND1 is a five-finger, 12-DoF hand with 120 three-axis force taxels per fingertip; net fingertip forces are transformed into the hand-base frame.Finger compliance control is used during experiments.Source: A. Experimental Setup · p. 5
  • Object Flipping requires a 90° reorientation ending in a stable grasp, while Screwdriver Rotation measures normalized progress toward 360° in 90° increments.Success rates are reported only for Object Flipping; screwdriver results use progress.Source: A. Experimental Setup · p. 5
  • The main evaluation uses 25 autonomous trials per object–method pair, pooling 50 trials for overall flipping and averaging object-specific means for overall screwdriver progress.Component and representation variants use 10 trials each, and their full-method results are separate from the main evaluation.Source: A. Experimental Setup · p. 5
  • For each skill, one frozen simulation-trained policy is reused across objects, while a separate object-specific policy is trained from correction data.Training uses 30 episodes per flipping object and 10 per screwdriver variant, selected from successful human-corrected executions.Source: A. Experimental Setup · p. 5
  • RMA directly executes the frozen base policy; Gated Correction and Continuous Correction learn residual actions from the same correction rollouts and reconstructed targets as ReDex.The residual baselines use base-policy targets and five-frame joint-position and fingertip-force histories; they are reference comparisons, not controlled ReDex ablations.Source: A. Experimental Setup · p. 5

V. Results

ReDex improves real-robot performance on both manipulation tasks across object variants, while using targeted finger guidance and retaining base-policy control of other fingers. Component studies show that tactile input, force-informed targets, and transition blending support autonomous flipping.

Real-world performance

  • Cuboid success rises from 4/25 (16%) to 20/25 (80%), and normal-screwdriver mean rotation progress rises from 26% to 95%.These results compare ReDex with the frozen RMA policy.Source: §V · pp. 5–6, Table I · p. 6

Adaptation across object variants

  • Across object variants, small- and large-screwdriver progress rises from 33% to 99% and 19% to 92%, while cellphone flipping success rises from 12% to 92%.The study uses one simulation-trained base policy per skill and object-specific policies trained from correction rollouts.Source: §V · p. 6, Table I · p. 6

Comparison with Correction Baselines

  • Gated and Continuous Correction reach 42% and 36% normal-screwdriver progress and 12% and 4% cuboid success, versus ReDex’s 95% and 80%.Both correction baselines remain below half of the target rotation on average and fail most flipping trials.Source: §V · p. 6, Table I · p. 6

Ablation Study

  • On cuboid flipping, success is 4/10 with motion supervision, 6/10 with tactile observations, and 9/10 with the complete method; removing force-informed targets or transition blending yields 7/10 or 6/10.Without tactile feedback the policy often stalls; force-informed targets preserve contact force information, and transition blending connects corrected segments with subsequent actions.Source: §V · pp. 6–7, Table II · p. 6

Tactile representations

  • Net-force observations outperform raw taxels with both MLP and GRU encoders, while MLP and GRU perform similarly with net-force input.The authors use net-force observations with an MLP encoder; they note raw taxels may be more sensitive to sensor noise and contact variation.Source: Table III · p. 6, §V · p. 7

Analysis of human effort

  • For cuboid flipping, ReDex uses 2.4 minutes of thumb guidance versus 13.2 minutes for full kinesthetic teaching, an 82% reduction; screwdriver guidance spans 38.8%–83.6% of rollout time.Cuboid guidance covers 24.0% of retained rollout time; screwdriver guidance involves one to three fingers, with the base policy controlling uncorrected fingers.Source: §V · p. 7, Table IV · p. 7, Figure 4 · p. 7

Failure analysis and limitations

  • Excessive or poorly timed interventions can disrupt coordination, and limited real-world trajectories and magnetic sensor drift may reduce robustness; evaluation covers two skills on one hardware platform.More complex tasks require further validation of joint coupling, feedback mechanisms, and intervention magnitude.Source: §V · p. 7

VI. Conclusion

ReDex combines finger-level physical correction with real-world tactile feedback to adapt simulation-trained policies. Across two skills and five objects, it improves real-world performance with limited correction data, without tactile simulation or additional simulation training.

VI. Conclusion

  • Across two skills and five objects, ReDex improves real-world performance with limited correction data, without tactile simulation or additional simulation training.Source: §VI · p. 8
  • During rollout, the base policy controls uncorrected fingers while the operator repairs local contact failures; reconstructed force-informed targets and transition blending supervise a standalone policy.Source: §VI · p. 8
  • Future work could test other tactile sensors, extend ReDex to wholebody manipulation, and investigate preserving simulation-learned adaptability during real-world correction training.Source: §VI · p. 8

Related papers

Loading 2610.07525v1…