PaperlayerView exact revision on arXiv

Source-linked AI summary

TacHair: Tactile Contact-Distribution Guided Online Correction for Robotic Hair Stroking and Perception

Ruiyi Hu, Yongqiang Zhao, Daniel Bak, Yupeng Wang, Xuyang Zhang, Shan Luo

arXiv:2610.10637v1cs.RO

TL;DR

Deformable hair can remain in contact while drifting across the end-effector, complicating regulation of the local interaction. TacHair uses spatial tactile contact distributions to guide residual corrections to a nominal visuotactile stroking policy; across 525 real-robot trials, it improved task success and contact maintenance over the same policy without correction.

  • Problem

    Deformable hair can remain in contact while shifting across the sensor, making spatially well-aligned interaction difficult to regulate and hard to observe beneath the end-effector.

    Source: §I · p. 1

  • Method

    TacHair represents tactile hair contact as a spatial distribution and uses a separately trained residual module to correct local deviations from a desired region alongside nominal visuotactile stroking.

    Source: §I · p. 2

  • Results

    Task success improved from 42.9% to 62.3% and contact maintenance from 59.4% to 88.0% versus the same visuotactile policy without online correction.

    Source: §I · p. 2

  • Takeaways & Limitations

    Spatial tactile contact distributions provide an intermediate representation between high-dimensional tactile sensing and reactive control for deformable, visually occluded interactions.

    Source: §I · p. 2

    The tested wigs vary jointly in multiple properties, and the main experiments use stationary headforms and a prescribed crown-to-back motion.

    Source: F. Discussion and Limitations · p. 7

Abstract

from arXiv · show

Hair stroking is common in daily grooming and personal care, and is also widely used in hair-product evaluation, motivating robots with similar physical interaction capabilities. Existing robotic hair-care and surface-following methods mainly rely on trajectory planning, compliance, force regulation, or tactile-conditioned policies, but deformable hair can remain in contact while gradually drifting across the end-effector, making local interaction difficult to regulate. We propose TacHair, a tactile contact-distribution guided online correction framework that represents high-resolution tactile observations as a spatial hair-contact distribution. A visuotactile imitation policy generates the nominal stroking motion, while a separately trained residual module corrects local contact deviations, separating task progression from contact recovery. We evaluate TacHair in 525 real-robot trials across five head geometries and three hair conditions. A successful stroke requires both sufficient task progression and contact maintenance; our method improves success from 42.9% to 62.3% and contact maintenance from 59.4% to 88.0% over the same visuotactile policy without correction. These results demonstrate spatial tactile contact distributions as an effective feedback representation for contact-preserving interaction with deformable and visually occluded surfaces. Demos, code, and datasets are available at https://tachair.github.io.

I. INTRODUCTION

Hair stroking requires continued task progression while regulating contact as deformable fibers shift beneath the end-effector. TacHair represents hair contact spatially and uses online residual correction to improve contact-preserving task success.

I. INTRODUCTION

  • Hair can remain in contact while its spatial location drifts across the sensor, a change external cameras struggle to observe beneath the end-effector.Trajectory progression, force regulation, or binary contact detection alone may not establish spatially well-aligned interaction.Source: §I · p. 1
  • TacHair represents hair–sensor contact as a spatial distribution, using a desired central region in tactile image space rather than reconstructing deformable hair’s instantaneous 3D contact geometry.The representation describes the location and extent of local contact evidence over the sensing surface.Source: §I · pp. 1–2
  • A nominal visuotactile policy handles stroke progression, while a separately trained residual module uses dedicated recentering demonstrations to recover local contact deviations.The correction module modifies nominal actions when the observed contact distribution leaves the desired region.Source: §I · p. 2
  • 62.3% task success and 88.0% contact maintenance, versus 42.9% and 59.4% without online correction, across 525 real-robot trials.Stroke-completion rate remained comparable; continuous contact metrics improved in 14 of 15 tested headform–hair combinations.Source: §I · p. 2

II. RELATED WORK

Prior robotic hair interaction work addresses configuration change, compliance, and force-aware manipulation, while tactile surface-following work regulates contact on structured surfaces. TacHair focuses on spatially aligned contact as hair deforms and shifts beneath the sensor.

II. RELATED WORK

  • Robotic hair-care systems have used compliant mechanisms, visual trajectory planning, and force feedback for washing, massage, combing, detangling, and other interactions.Examples include autonomous combing, force-adaptive detangling, and a compliant end-effector for contact-rich hair-care tasks.Source: A. Robotic Hair Interaction and… · p. 2
  • TacHair targets spatially aligned contact during continuous stroking, where deforming hair can remain in contact while its distribution shifts across the sensor.This focus differs from prior work addressing hair configuration, compliant interaction, or force-aware manipulation.Source: A. Robotic Hair Interaction and… · p. 2
  • Tactile surface-following methods use contact pose, shear, or contour features for closed-loop regulation, typically on rigid or geometrically structured interfaces,,.These methods demonstrate tactile feedback for contact regulation, while hair creates a deformable, changing interaction layer.Source: A. Robotic Hair Interaction and… · p. 2

B. Visuotactile Policy Learning and Online Tactile Correction

The paper distinguishes explicit spatial contact-distribution feedback from approaches that use tactile sensing for policy generation or reactive refinement. Its control formulation assigns stroke progression to a nominal visuotactile policy and local contact recovery to a residual module.

B. Visuotactile Policy Learning and Online Tactile Correction

  • Visuotactile policies commonly incorporate tactile observations for action generation, while reactive methods refine nominal actions; TacHair distinguishes itself through explicit spatial contact-distribution feedback.Deviation from a desired region activates and conditions learned residual corrections during stroking.Source: B. Visuotactile Policy Learning and… · pp. 2–3

A. Task Formulation and System Overview

  • The nominal policy drives crown-to-back motion, and a residual joint action recenters contact when its estimated tactile centroid leaves the predefined central region.The final command combines nominal and residual actions under shared joint limits, displacement constraints, and temporal smoothing.Source: A. Task Formulation and System Overview · p. 3
  • A successful trial must satisfy both stroke completion and contact maintenance.The two requirements separate task progression from maintaining the desired sensor–hair interaction.Source: A. Task Formulation and System Overview · p. 3

B. Tactile Hair-Contact Distribution

TacHair converts high-resolution tactile observations into a compact spatial distribution of hair-contact evidence. The distribution is normalized and summarized with statistics used for online correction.

B. Tactile Hair-Contact Distribution

  • Aligned tactile frames are encoded into a spatial feature grid, and a lightweight decoder predicts hair-contact evidence across 20 lateral bands.The representation operates over tactile image evidence rather than reconstructed metric 3D contact points or forces.Source: B. Tactile Hair-Contact Distribution · p. 3
  • Manually annotated tactile contact bands supervise the decoder, which uses class-weighted binary cross-entropy while the tactile encoder remains frozen.Each band output represents the probability of visible hair-contact traces in that band.Source: B. Tactile Hair-Contact Distribution · p. 3
  • After normalization, the spatial distribution yields regional masses, spread, entropy, and a normalized lateral centroid for online correction.The centroid summarizes the distribution’s lateral location on the tactile surface.Source: B. Tactile Hair-Contact Distribution · p. 3

C. Nominal Visuotactile Stroking Policy

The nominal visuotactile policy drives the overall stroke, while a separately trained residual module uses tactile contact distributions to recover local contact deviations.

C. Nominal Visuotactile Stroking Policy

  • tactile-SmolVLA generates the crown-to-back stroke from RGB, tactile, robot-state, and task-instruction inputs, predicting sequences of absolute joint-position targets.Its tactile frames use a frozen Sparsh encoder and a lightweight cross-attention adapter.Source: C. Nominal Visuotactile Stroking Policy · p. 3

D. Contact-Distribution Guided Online Correction

  • Dedicated recentering demonstrations train the correction module to produce short-horizon joint-space residuals when contact leaves the desired central tactile region.The expert target is a joint-position command approximately 0.5 seconds ahead.Source: D. Contact-Distribution Guided Online… · p. 3
  • The framework adds a contact-conditioned residual to the nominal joint target, separating overall stroke progression from local tactile contact regulation.The correction is conditioned on the contact distribution and current robot state.Source: Figure 2 · p. 4
  • Two correction branches handle centroid-increasing and centroid-decreasing transitions, with the selected residual activated by contact deviation.Contact within the desired region leaves the nominal action unchanged; applied corrections are scaled, clipped, constrained, and temporally smoothed.Source: D. Contact-Distribution Guided Online… · pp. 3–4

E. Evaluation Metrics

The evaluation treats stroke progression and contact preservation as separate requirements for success, and measures contact using image-based tactile proxies.

E. Evaluation Metrics

  • Contact quality combines central-region coverage, which penalizes lateral contact loss, with the longest contiguous active span intersecting that region.The evaluation also reports a time-weighted qualified-contact fraction, mean contact quality, and Tail-Q.Source: E. Evaluation Metrics · p. 4
  • Normalized stroke progress is measured in five-dimensional arm joint space, excluding the gripper.The initial, current, and terminal arm configurations define the progress measure.Source: E. Evaluation Metrics · p. 4
  • A trial succeeds only when it meets both the stroke-completion and contact-maintenance criteria.The evaluation reports progression and contact preservation separately before combining them for trial-level success.Source: E. Evaluation Metrics · p. 4
  • The tactile metrics are image-based proxies, not measurements of physical force or comfort.The controller and evaluator share a frozen tactile representation, but contact quality is computed offline and is not directly optimized by the correction controller.Source: E. Evaluation Metrics · p. 4

A. Experimental Setup and Training Protocol

The setup uses separate SO101 and UR5e policies and evaluates hair stroking across five headforms and three joint hair conditions. Training separates nominal stroking, contact prediction, and recentering data.

A. Experimental Setup and Training Protocol

  • The SO101 is the primary quantitative platform, while the UR5e provides additional cross-platform validation with separately trained policies and correction modules.The robots have different kinematics and joint spaces.Source: A. Experimental Setup and Training… · p. 5
  • Training uses 97 nominal demonstrations, 1,045 manually annotated tactile images, and 75 dedicated recentering demonstrations, with the modules trained separately.The nominal dataset contains 7,002 frames; training data use the Small headform with Hair A.Source: A. Experimental Setup and Training… · p. 5
  • Evaluation covers five standardized CDC/NIOSH headforms and three hair conditions that differ jointly in length, curl, and smoothness.The headforms are shown at original scale; the hair-condition illustrations are qualitative schematics, not photographs or physical measurements.Source: A. Experimental Setup and Training… · p. 5, Figure 4 · p. 5, Figure 5 · p. 5

B. Compared Methods and Evaluation Protocol

The evaluation compares tactile-conditioned stroking with and without online correction across 525 trials and 15 headform–hair combinations. Correction improves task success and contact metrics, with a qualitative UR5e demonstration complementing the SO101 results.

B. Compared Methods and Evaluation Protocol

  • The three configurations are vision-only SmolVLA, tactile-conditioned Tactile-SmolVLA, and Ours, which adds correction to the same tactile-policy checkpoint.The Tactile-SmolVLA–Ours comparison isolates explicit online correction.Source: B. Compared Methods and Evaluation… · p. 5
  • The main evaluation uses 525 SO101 trials across 15 combinations, with 175 trials per method; each trial starts with contact near the sensor center and proceeds without physical guidance.Within a common 45-second window, task success requires both stroke completion and contact maintenance, using fixed thresholds across methods.Source: B. Compared Methods and Evaluation… · pp. 5–6

C. Results Across Headforms and Hair Conditions

  • All three continuous contact metrics improve in 14 of 15 combinations; mean contact quality rises 0.459 to 0.649, FQ(0.60) 0.422 to 0.635, and Tail-Q 0.480 to 0.746.Small–Hair-C is the exception, with decreases of 0.237, 0.319, and 0.139 in those metrics, respectively; the passage attributes this pattern to the combined local geometry and hair condition.Source: Figure 6 · p. 6, C. Results Across Headforms and Hair… · p. 6
  • 62.3% task success versus 42.9% for Tactile-SmolVLA and 36.6% for SmolVLA; contact maintenance rises from 59.4% to 88.0%, with comparable completion (74.3% versus 72.6%).The task-success improvement over Tactile-SmolVLA is 19.4 percentage points (95% bootstrap confidence interval [11.4, 27.4]).Source: Table I · p. 6, C. Results Across Headforms and Hair… · p. 6
  • Among 401 episodes with complete observation windows, 145 reach the contact target within 8 seconds, with a median confirmation time of 4.432 seconds.Figure 7 also compares contact quality and time-weighted lateral error before correction and after target attainment for 70 confirmed episodes.Source: Figure 7 · p. 6
  • The UR5e sequence qualitatively shows continued stroke progression with online contact recentering, relating global robot motion to local hair–sensor interaction.The validation uses a separately trained UR5e policy and correction module, so it tests framework applicability rather than zero-shot transfer.Source: Figure 8 · p. 7, C. Results Across Headforms and Hair… · p. 7

D. Analysis of Online Tactile Correction

Episode-level analysis measures target attainment after residual correction and compares contact quality and lateral error before and after confirmed attainment. The reported pre/post comparison is limited to episodes that reached the target.

D. Analysis of Online Tactile Correction

  • 145 of 401 episodes with complete 8-second observation windows reach the contact target (36.2%), with a median confirmation time of 4.432 seconds.The target requires valid tactile contact, Qt ≥0.60, and instantaneous lateral error et ≤0.08 for three consecutive observations.Source: D. Analysis of Online Tactile Correction · p. 7
  • For 70 confirmed episodes, mean contact quality rises from 0.678 to 0.883 while time-weighted mean lateral error falls from 0.099 to 0.049.The comparison uses complete 2-second windows before correction onset and after confirmation.Source: D. Analysis of Online Tactile Correction · p. 7
  • Failure to reach the stricter target does not imply no incremental benefit; trial-level results remain the primary measure of overall performance.The target-attainment criterion is stricter than detecting incremental improvement.Source: D. Analysis of Online Tactile Correction · p. 7

E. Qualitative Stroking Sequence

The qualitative sequence illustrates contact recentering during continued stroking, while the discussion bounds the evidence to stationary headforms and a prescribed stroke. The conclusion reports improved contact-preserving performance and identifies broader evaluation as future work.

E. Qualitative Stroking Sequence

  • The UR5e sequence qualitatively links global robot motion to local hair–sensor interaction as contact is recentered while stroke progression continues.It uses a robot embodiment and sensor mounting different from the primary SO101 platform.Source: E. Qualitative Stroking Sequence · p. 7

F. Discussion and Limitations

  • The main experiments use stationary headforms and a prescribed crown-to-back stroke; tested wigs vary jointly in multiple properties.Moving human heads, additional stroking directions, and independent measurements of contact force and human comfort remain future evaluation directions.Source: F. Discussion and Limitations · p. 7
  • The reported contact metrics share the controller’s frozen tactile representation, and the contact-quality score is not directly optimized during control.The conclusion identifies adaptive correction, moving heads, additional stroke directions, and independent contact-force and comfort measurements as future work.Source: F. Discussion and Limitations · p. 7, §V · p. 8

V. CONCLUSION

  • Across 525 SO101 trials, task success rises from 42.9% to 62.3% and contact maintenance from 59.4% to 88.0% versus the same policy without correction.Continuous contact metrics improve in 14 of 15 tested combinations; qualitative validation also uses a UR5e and an alternative tactile membrane.Source: §V · p. 8
Loading 2610.10637v1…