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

arXiv:2610.10637 · cs.RO · Submitted 2026-10-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "TacHair: Tactile Contact-Distribution Guided Online Correction for Robotic Hair Stroking and Perception".

Dev: The gist:

Rosa: First, who's behind it and why it matters.

Paper summary: Rosa: So, we're looking at this paper called TacHair, which tackles robotic hair stroking by proposing a way to handle that tricky bit where the hair keeps moving under the robot's contact.

Dev: Yeah, it’s about using this tactile information not just for basic control, but specifically to track where the actual contact is happening across the surface.

Taro: It seems like they are trying to separate what we call task progression from what we call contact recovery, which is a big hurdle when dealing with something as floppy as hair one <ref:2610.10637#pg3>.

Rosa: Exactly. The core idea of TacHair is that instead of just looking at the overall motion, the system explicitly represents local hair-sensor interaction as this spatial tactile contact distribution.

Dev: So it’s not just reacting to a force reading; it’s mapping those high-resolution tactile observations into a spatial feature grid to get a picture of where the hair is actually touching.

Taro: I wonder if that's helpful because when hair drifts, the whole system gets confused about whether it should keep moving forward or try to re-establish contact one <ref:2610.10637#pg3>.

Rosa: That’s the problem they’re solving. They use a visuotactile imitation policy for the overall motion, which gives you that nominal stroking path, and then a separate residual module fixes any deviations from where the hair is supposed to be touching.

Dev: And they achieve this by defining the desired interaction directly in tactile image space as a central contact region, so they don't have to do all that complicated three dee reconstruction of the hair geometry one.

Taro: That sounds smart because reconstructing three dee geometry of deformable hair in real-time is usually a huge computational headache for any robot system one.

Rosa: Right. And when they look at what actually works, they used five hundred twenty-five real-robot trials across five different head shapes and three different hair conditions to test this framework <ref:2610.10637#pg1,525 real-robot trials across five>.

Dev: The results show that this approach improves task success from forty-two point nine percent up to sixty-two point three percent, but more importantly, it boosts contact maintenance from nearly sixty percent up to eighty-eight percent one.

Taro: That jump in contact maintenance is significant because maintaining that physical connection while moving across a deforming surface is genuinely hard one <ref:2610.10637#pg3>.

Rosa: It shows that having this explicit spatial representation of the hair interaction gives the robot much better feedback for preserving contact during the stroke.

Dev: The paper argues that this explicit representation, which they call a compact tactile contact distribution, acts as an effective intermediate step between raw tactile sensing and what the control system needs to react one <ref:2610.10637#pg3>.

Taro: So, what does this mean for a robot that’s supposed to do something delicate like grooming? Does it mean it can handle real-world mess better?

Rosa: It means the robot isn't just following a path; it’s actively monitoring the contact distribution and making immediate local adjustments when things drift away from the center, which is what they call online correction.

Dev: The system takes that distribution, plus some statistics about it and the robot's current joint state, to predict a residual correction action using a shared neural network one <ref:2610.10637#pg3>.

Taro: I’m interested in the mechanism of that prediction. How does the AI figure out what change is needed when something goes sideways?

Rosa: They use a shared MLP with two output heads trained with a Smooth L1 loss to predict those residual joint actions based on the input contact distribution data one <ref:2610.10637#pg3>.

Dev: The final command they send is just adding that predicted residual action to the nominal action, but then it runs through some limits and smoothing filters before being executed one <ref:2610.10637#pg3>.

Taro: So, if we look at what this means for autonomy research, it suggests that instead of needing a perfect model of the hair's physics, we can use localized tactile feedback to manage uncertainty in real-time one <ref:2610.10637#pg3>.

Rosa: Precisely. It moves the focus from trying to perfectly model the entire complex deformation to focusing on correcting local contact errors guided by that spatial distribution.

Dev: It’s a way of decoupling the high-level movement from the fine-grained, moment-to-moment contact stability, which is crucial for those continuous grooming tasks one <ref:2610.10637#pg3>.

Taro: And looking at the authors, Hu, Zhao, Bak, Wang, Zhang and Luo all from King's College London; it’s a solid team working on practical applications of this kind of vision and tactile integration one <ref:2610.10637#pg3>.

Rosa: So when we think about the title "TacHair: Tactile Contact-Distribution Guided Online Correction for Robotic Hair Stroking and Perception," it really captures the essence of what they built here.

Dev: It’s not just about stroking anymore; it's about using that spatial contact distribution to guide the correction module, which is a key part of their contribution one <ref:2610.10637#pg3>.

Taro: For someone listening who doesn't know robotics, it boils down to this: they gave the robot a way to "feel" where its fingers are touching the hair and use that feel instantly to fix any slippage during the stroke one <ref:2610.10637#pg3>.

Rosa: That’s a good way to put it. It shows that making robots interact with deformable surfaces is possible when you provide them with this kind of detailed, spatial feedback loop one <ref:2610.10637#pg3>.

Conclusion: Rosa: So we’ve been looking at TacHair, this paper by Hu and his team about robotic hair stroking and perception. It boils down to using tactile data to fix mistakes while the robot is moving.

Dev: Yeah, it’s about that contact distribution thing they introduce, mapping the raw touch into a spatial map so the system knows exactly where the hair is touching at any given moment.

Taro: I see what you mean. It separates the main task—the stroking motion—from the recovery part, which is usually really tricky when dealing with something floppy like hair.

Rosa: Exactly. They show that this explicit contact distribution representation helps the robot maintain that connection much better during a stroke compared to just using standard policies without any correction loop.

Dev: And they got some pretty solid numbers there, improving task success from about forty-three percent up to sixty-two percent in their real-world tests. That’s a tangible gain for any control engineer looking at reliability.

Taro: But what this really changes is how we think about feedback for these kinds of interactions. It suggests that you don't need a perfect three dee model of the hair to get good results; local tactile statistics are actually a useful way to manage uncertainty in real-time.

Rosa: That’s the big picture, right? It moves us away from relying solely on pre-programmed trajectories and toward systems that can actively sense and correct their physical interaction with deformable surfaces.

Dev: From an engineering standpoint, the latency of that correction loop matters a lot, so they show how they integrate it without completely bogging down the control rate.

Taro: So if you look at the conclusion of this paper, it really says that this contact distribution approach is a way to complement tactile conditioning in the nominal policy. It’s not replacing it; it’s giving that policy a better way to handle those moments when things go sideways.

Rosa: Right. And they found that this explicit representation complements the conditioning and substantially improves contact maintenance without changing how well the robot completes its stroke overall.

Dev: So what does this mean for future work? The authors mention looking at adaptive correction strategies and moving heads, which suggests they know there are still a lot to explore outside of these specific trials.

Taro: Yeah, they also mentioned that future work could involve getting independent measurements of contact force and how comfortable the human feels during these interactions. That would be a big step toward making these robots truly useful in real-world grooming scenarios.

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

King's College London

cs.RO

Submitted: 2026-10-07

Updated: 2026-10-07

Code: https://github.com/TheRobotStudio/SO-ARM100

Project page: https://tachair.github.io

The gist: The gist: TacHair proposes a tactile contact-distribution guided online correction framework for robotic hair stroking that improves task success and contact maintenance by explicitly representing

Key concepts

Tactile Contact Distribution
This concept maps high-resolution tactile sensor data into a spatial representation showing where the robot is actually touching the hair. It uses an encoder and decoder to predict contact evidence across several bands, creating a spatial map that summarizes the local interaction for control.
Nominal Stroking Policy
This is the primary policy trained from expert demonstrations that generates the overall motion of stroking from one end of the head to another. It handles the general path planning and movement under expected, near-nominal contact conditions.
Contact Recovery Module
A separately trained module that acts as a correction mechanism. It is activated only when the tactile data shows that the actual contact distribution has moved outside a predefined central target area. It predicts small adjustments to the robot's joint actions to bring the interaction back into place.

Terminology

Summary

The gist: TacHair proposes a tactile contact-distribution guided online correction framework for robotic hair stroking that improves task success and contact maintenance by explicitly representing local hair–sensor interaction as a spatial tactile contact distribution.

How it works

The proposed framework separates the control problem into nominal stroking and contact recovery, where the nominal visuotactile policy generates the overall motion, while a separately trained residual module corrects local deviations<ref:2610.10637#pg3> The desired interaction is defined directly in tactile image space as a central contact region, avoiding the need to reconstruct the instantaneous 3D contact geometry of deformable hair<ref:2610.10637#pg3> The nominal policy is learned from complete expert demonstrations and generates the overall crown-to-back motion under near-nominal interaction<ref:2610.10637#pg3> Because substantial contact deviations and recovery behaviors are sparse in such demonstrations, a residual module is trained to modify the nominal action only when the observed contact distribution leaves the desired region<ref:2610.10637#pg3>

Tactile Hair-Contact Distribution

The paper represents local hair–sensor interaction by mapping high-resolution tactile observations to a spatial contact distribution<ref:2610.10637#pg4> This involves encoding the current and preceding high-resolution tactile frames into a spatial feature grid Ft using a frozen Sparsh-DINO encoder<ref:2610.10637#pg4> A lightweight decoder then predicts hair-contact evidence over K = 20 uniformly spaced bands along the lateral image axis, resulting in pt,i which denotes the probability of visible hair-contact traces in band i<ref:2610.10637#pg5> The normalized contact distribution wt describes the relative spatial contact distribution and µt is its normalized lateral centroid<ref:2610.10637#pg5> This representation includes compact distribution statistics such as regional masses, spread, entropy, and the centroid for online correction<ref:2610.10637#pg5>

Contact-Distribution Guided Online Correction

The correction module is activated when the estimated contact moves outside a predefined central region on the tactile surface<ref:2610.10637#pg5> The correction input xt contains the normalized contact distribution, six distribution statistics, and the six-dimensional robot joint state<ref:2610.10637#pg5> A shared MLP with two six-dimensional output heads predicts the corresponding residual correction and is trained using a Smooth L1 loss<ref:2610.10637#pg5> The final command is given by at = S (a nom t + ∆at), where ∆at is the predicted residual joint action, and S applies shared joint limits, displacement constraints, and temporal smoothing<ref:2610.10637#pg5>

Evaluation and Results

The framework was evaluated in 525 real-robot trials across five head geometries and three hair conditions<ref:2610.10637#pg3> A successful stroke requires both sufficient task progression and contact maintenance<ref:2610.10637#pg5> Compared with the same visuotactile policy without online correction, the method improves task success from 42.9% to 62.3% and contact maintenance from 59.4% to 88.0%, while maintaining a comparable strokecompletion rate<ref:2610.10637#pg5> These results demonstrate that explicit contact recovery complements tactile conditioning within the nominal policy and that spatial tactile contact distributions provide a useful intermediate representation between high-dimensional tactile sensing and reactive control for deformable, visually occluded interactions<ref:2610.10637#pg3>

Conclusion

The main contributions of this work are formulating robotic hair stroking as a deformable surface-following problem and introducing a compact tactile contact distribution to explicitly represent the location and extent of local hair–sensor interaction<ref:2610.10637#pg6> The results indicate that explicit contact-distribution guided recovery complements tactile conditioning in the nominal policy, substantially improving contact maintenance without materially changing stroke completion Future work will investigate adaptive correction strategies, moving heads and additional stroking directions, and independent measurements of contact force and human comfort

--- Page 1 ---

TacHair: Tactile Contact-Distribution Guided Online Correction for Robotic Hair Stroking and Perception<ref:2610.10637#pg2> Ruiyi Hu, Yongqiang Zhao, Daniel Bak, Yupeng Wang, Xuyang Zhang, Shan Luo are with King’s College London, London WC2R 2LS, United Kingdom. Email: yongqiang.zhao@kcl.ac.uk.<ref:2610.10637#pg2> Abstract— 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<ref:2610.10637#pg2> Existing robotic hair-care and surface-following methods mainly rely on trajectory planning, compliance, force regulation, or tactile-conditioned policies<ref:2610.10637#pg2> But deformable hair can remain in contact while gradually drifting across the end-effector, making local interaction difficult to regulate<ref:2610.10637#pg2> We propose TacHair, a tactile contact-distribution guided online correction framework that represents high-resolution tactile observations as a spatial hair-contact distribution<ref:2610.10637#pg3> 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<ref:2610.10637#pg3> We evaluate TacHair in 525 real-robot trials across five head geometries and three hair conditions<ref:2610.10637#pg5> 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<ref:2610.10637#pg5> These results demonstrate spatial tactile contact distributions as an effective feedback representation for contact-preserving interaction with deformable and visually occluded surfaces<ref:2610.10637#pg5> I. INTRODUCTION Hair stroking is a common physical interaction in daily grooming, personal care, and interpersonal assistance<ref:2610.10637#pg6> Similar stroking and combing motions are also widely used in hair-product evaluation and hair-condition assessment, where the interaction between an applicator and hair provides information about properties such as smoothness, friction, and combability [1], [2]<ref:2610.10637#pg6> As robots increasingly enter domestic, assistive, and product-development environments, it is therefore desirable for them to acquire comparable capabilities for sustained and adaptive physical interaction [3], [4]<ref:2610.10637#pg6> Unlike conventional pick-and-place manipulation, hair stroking requires the robot to continuously progress along a curved surface while maintaining an appropriate local interaction with highly deformable fibers<ref:2610.10637#pg6> The difficulty lies not only in reaching the desired region, but also in regulating contact throughout the entire motion as the interaction interface evolves<ref:2610.10637#pg6> Existing robotic hair-care systems address related problems mainly through visual trajectory planning, compliant mechanisms, and force feedback<ref:2610.10637#pg6> Prior work has demonstrated autonomous hair combing using hair segmentation and path planning [5], adaptive brushing based on interaction forces [6], compliant head-care systems for washing and massage [7], and soft end-effectors for contact-rich tasks such as head patting and finger combing [8]<ref:2610.10637#pg6> In parallel, tactile surface-following methods estimate contact pose, shear, or contour features and use them for closed-loop servoing over rigid or geometrically structured surfaces [9], [10], [11]<ref:2610.10637#pg6> These methods provide important mechanisms for maintaining physical contact, but hair stroking introduces a different interaction regime<ref:2610.10637#pg6> As hair fibers deform, slide, and redistribute beneath the end-effector, contact may remain present while its spatial location gradually shifts across the tactile surface<ref:2610.10637#pg6> This drift is particularly difficult to observe using an external camera because the relevant interaction occurs directly underneath the end-effector<ref:2610.

Improvements for AI systems

  1. A contact-distribution guided online correction framework that separates task progression from contact recovery allows robots to maintain spatial alignment during stroking by using spatial tactile contact distributions as an effective feedback representation for contact-preserving interaction with deformable and visually occluded surfaces. This enables the system to achieve a contact maintenance from 59.4% to 88.0% compared to baselines, even when hair drifts across the end-effector.

  2. The framework improves task success by incorporating explicit correction, as it improves success from 42.9% to 62.3% and achieves a contact maintenance from 59.4% to 88.0% over the same visuotactile policy without correction, demonstrating that spatial tactile contact distributions provide a useful intermediate representation between high-dimensional tactile sensing and reactive control.

  3. The system can perform robust contact recovery by using a contact-distribution guided online residual correction framework, where the correction module is trained on dedicated recentering demonstrations from laterally biased contacts to modify the nominal action only when the observed contact distribution leaves the desired region.

  4. The robot can dynamically adjust its local interaction based on tactile feedback; specifically, if the estimated contact moves outside a predefined central region, the system generates a residual joint action to recenter the interaction, ensuring that contact within the desired central region leaves the nominal action unchanged.

  5. The AI system can generate high-resolution spatial feature grids from tactile images using a frozen Sparsh-DINO encoder and a lightweight decoder to predict hair-contact evidence over bands, allowing it to estimate metrics like the normalized lateral centroid µt for online control.

Sources

Related papers