PaperlayerView exact revision on arXiv

Source-linked AI summary

DYMO-Hair: Generalizable Volumetric Dynamics Modeling for Robot Hair Manipulation

Chengyang Zhao, Uksang Yoo, Arkadeep Narayan Chaudhury, Giljoo Nam, Jonathan Francis, Jeffrey Ichnowski, Jean Oh

arXiv:2510.06199v2cs.RO

TL;DR

Hair’s fine structure and complex dynamics make robot hair care challenging, while existing systems have limited generalization across hairstyles and goals. DYMO-Hair combines volumetric dynamics learning with a model-based planner for visual goal-conditioned styling. It reports improved closed-loop results on unseen hairstyles in simulation and zero-shot transfer to wigs in real-world demonstrations.

  • Problem

    Hair’s fine structure and complex dynamics challenge robot hair care, and existing approaches restrict generalization across diverse hairstyles and goals.

    Source: §I · p. 2

  • Method

    DYMO-Hair learns volumetric hair dynamics through action-conditioned latent state editing and diverse synthetic pretraining, then uses an MPPI-based planner for visual goal-conditioned styling.

    Source: §I · p. 2

  • Results

    22% lower final geometric error and 42% higher absolute success rate than the state-of-the-art system are reported for closed-loop styling; real-world demonstrations show zero-shot transfer to wigs.

    Source: §I · p. 2

  • Takeaways & Limitations

    The demonstrations show model-based visual goal-conditioned hair styling on unseen hairstyles in simulation and with zero-shot transfer to physical wigs.

    Source: §I · p. 2

Abstract

from arXiv · show

Hair care is an essential daily activity, yet it remains inaccessible to individuals with limited mobility and challenging for autonomous robot systems due to the fine-grained physical structure and complex dynamics of hair. In this work, we present DYMO-Hair, a model-based robot hair care system. We introduce a novel dynamics learning paradigm that is suited for volumetric quantities such as hair, relying on an action-conditioned latent state editing mechanism, coupled with a compact 3D latent space of diverse hairstyles to improve generalizability. This latent space is pre-trained at scale using a novel hair physics simulator, enabling generalization across previously unseen hairstyles. Using the dynamics model with a Model Predictive Path Integral (MPPI) planner, DYMO-Hair is able to perform visual goal-conditioned hair styling. Experiments in simulation demonstrate that DYMO-Hair's dynamics model outperforms baselines on capturing local deformation for diverse, unseen hairstyles. DYMO-Hair further outperforms baselines in closed-loop hair styling tasks on unseen hairstyles, with an average of 22% lower final geometric error and 42% higher success rate than the state-of-the-art system. Real-world experiments exhibit zero-shot transferability of our system to wigs, achieving consistent success on challenging unseen hairstyles where the state-of-the-art system fails. Together, these results introduce a foundation for model-based robot hair care, advancing toward more generalizable, flexible, and accessible robot hair styling in unconstrained physical environments. More details are available on our project page: https://dymohair.github.io/.

I. INTRODUCTION

DYMO-Hair addresses limited generalization in robot hair care with a volumetric dynamics model and a model-based visual goal-conditioned styling system. Simulation and real-world demonstrations report stronger performance on unseen hairstyles.

I. INTRODUCTION

  • Existing robot hair-care systems rely on handcrafted trajectories or rule-based controllers, restricting generalization across diverse hairstyles and goals.Hair also poses representation, structure, supervision, and data challenges, including slow real-world collection and difficult resets across styles.Source: §I · pp. 1–2
  • The dynamics model represents hair with high-resolution volumetric occupancy and orientation, then uses action-conditioned latent state editing for generalizable local deformation modeling.A compact 3D latent space is pretrained on diverse synthetic hair states; dense occupancy and orientation supervision supplies local signals for fine-grained deformations.Source: §I · pp. 1–2
  • 22% lower final geometric error and 42% higher absolute success rate than the state-of-the-art system are reported for closed-loop styling.Real-world demonstrations also show zero-shot transfer to wigs, with consistent success on challenging unseen hairstyles where the state-of-the-art system fails.Source: §I · p. 2
  • The work studies model-based robot hair manipulation and presents DYMO-Hair, a visual goal-conditioned styling system evaluated across diverse hairstyles in simulation and real-world settings.Its MPPI-based planner uses the dynamics model for closed-loop styling.Source: Figure 1 · p. 1, §I · p. 2
  • A novel position-based dynamics method simulates strand-level, contact-rich hair combing to generate large-scale synthetic data across diverse hairstyles.The simulator supports model pretraining and closed-loop experiments.Source: §I · p. 2

II. RELATED WORKS

Prior deformable-object models face limits in fine-grained hair representation, scalability, and dense supervision, while existing robot hair systems often lack flexible goal-driven control. The paper frames hair manipulation as 3D state estimation and dynamics modeling, using volumetric states to support generalizable manipulation.

B. Robot System for Hair Manipulation

  • Prior robot hair systems use constrained feedback or flow-aligned combing, while 2D rule-based goal planning remains restricted in hairstyle and goal scope.The proposed direction captures 3D states, models 3D dynamics, and targets generalizable, flexible, goal-conditioned manipulation.Source: B. Robot System for Hair Manipulation · p. 2

C. Hair Dynamics Simulation

  • Hair dynamics methods trade computational efficiency against strand-level fidelity, while learning-based approaches still struggle with fine detail, scalability, or dense supervision.Clump-level models are efficient but miss strand behavior; strand-based physical models are costly, and learning-based methods have not achieved strand-level, physically accurate simulation.Source: C. Hair Dynamics Simulation · p. 2, A. Dynamics Modeling for Deformable… · p. 2
  • The paper proposes a PBD simulator for efficient, visually realistic, physically plausible contact-rich hair-combing dynamics and synthetic-data generation.The simulator also serves as a testbed for closed-loop hair-manipulation experiments.Source: C. Hair Dynamics Simulation · p. 3

III. PROBLEM FORMULATION

  • The system iteratively estimates hair state from observations and selects 3D combing motions to advance toward a user-specified visual goal.The predicted next state is given by the dynamics model under the selected action.Source: §III · p. 3

A. State Representation

  • Hair is represented as a high-resolution volumetric occupancy grid with a 3D orientation field, retaining geometric and structural detail while enabling dense voxel-level supervision.The voxelized representation balances computational efficiency and geometric fidelity when resolution is sufficiently high.Source: A. State Representation · p. 3

B. State Latent Space Pre-training

DYMO-Hair pre-trains a compact volumetric latent space from diverse hairstyles, then edits that space using encoded combing actions to model dynamics. Its supervision combines occupancy, orientation, and codebook losses to support geometric, directional, and latent representation quality.

B. State Latent Space Pre-training

  • A two-phase ControlNet-style paradigm targets generalizable hair dynamics, with Phase 1 pre-training a hierarchical 3D latent space on diverse hairstyles and combing deformations.Hierarchical quantization captures complementary local and global information for state modeling and reconstruction.Source: B. State Latent Space Pre-training · pp. 3–4

C. Dynamics: Action-conditioned State Editing

  • In Phase 2, a trainable motion path encodes combing actions and fuses them with frozen pre-trained state features to edit latent embeddings and decode the next state.The motion is voxelized into a time-indexed distance map around the tool path to align control with the state spatially.Source: C. Dynamics: Action-conditioned State… · p. 4

D. Supervision

  • The composite loss combines focal and soft Dice occupancy losses, orientation L1 loss with directional symmetry, and EMA commitment loss to train geometry, direction, and codebook matching.The occupancy losses address class imbalance and encourage volumetric overlap with ground truth.Source: D. Supervision · p. 4

V. HAIR DYNAMICS SIMULATION

The hair simulator uses synthetic data and a particle-based PBD model to generate contact-rich combing dynamics across diverse hairstyles. Three strand constraints approximate physical behavior, with twist handled heuristically to balance accuracy and computational efficiency.

V. HAIR DYNAMICS SIMULATION

  • Large-scale diverse hair dynamics data is impractical to collect in the real world, so the authors use a GPU-accelerated synthetic simulator for visually realistic, physically plausible strand-level combing.Messy, hard-to-reset entangled hair states make real-world data collection time-consuming.Source: §V · p. 4
  • The PBD simulator models each strand as a particle sequence rooted at a fixed scalp point and applies stretch, bending, and twist constraints.Thousands of strands are simulated in parallel, and particle positions are iteratively updated toward satisfying all constraints.Source: §V · pp. 4–5
  • Twist is approximated with skip-connected 2D bending constraints across intersecting planes, while standard PBD collision and friction handling model strand-tool and inter-strand contacts.The twist heuristic balances accuracy and computation efficiency against more expensive physically detailed approaches.Source: §V · p. 5
  • The simulator provides efficient hair-combing dynamics across diverse hairstyles for model-learning data and closed-loop system experiments.Its intended scope includes both synthetic-data generation and evaluation of the hair-styling system.Source: §V · p. 5

VI. MODEL-BASED ROBOT HAIR STYLING SYSTEM

DYMO-Hair uses an MPPI-based model-predictive-control loop to predict hair states and optimize actions toward a visual goal. Its dynamics model is evaluated on unseen hairstyles, with experiments also describing the training data and physical test setup.

VI. MODEL-BASED ROBOT HAIR STYLING SYSTEM

  • The MPPI-based planner rolls out candidate actions through the dynamics model and optimizes a trajectory to reduce geometric distance from predicted hair states to the goal.The system executes an action chunk, updates its observation, and repeats until the goal is reached.Source: §VI · p. 5

A. Experiment Setup

  • Ten hairstyles train the dynamics model, while seven unseen simulation hairstyles and two physical wigs are reserved for evaluation.The simulation hairstyles and mannequin head come from USC-HairSalon.Source: A. Experiment Setup · p. 5, B. Hair Combing Dynamics Learning · p. 5
  • The real-world setup uses a calibrated robot arm, a wrist-mounted RGB-D camera, and a 3D-printed tool with a deformable fingertip.The tool is designed to resemble the simulated cylinder and reduce the sim-to-real gap.Source: A. Experiment Setup · p. 5

B. Hair Combing Dynamics Learning

  • On unseen hairstyles, DYMO-Hair’s dynamics model outperforms the baselines in capturing local hair deformation near the comb.The evaluation reports mean and 90th-percentile performance to capture overall behavior and sparse, significant deformations.Source: Table I · p. 6, B. Hair Combing Dynamics Learning · p. 6

C. Closed-loop Goal-conditioned Hair Styling

DYMO-Hair combines a learned hair dynamics model with MPPI planning for closed-loop, visual goal-conditioned styling. It outperforms comparison systems in simulation and transfers to physical wigs, while the authors identify assumptions and interaction constraints that limit current real-world usability.

VIII. CONCLUSION

  • The system assumes geometrically aligned goal configurations and, for real-world use, privileged mannequin-head information and perfect calibration.The authors identify online head estimation and action-space constraints near safety-critical regions such as the eyes as future needs.Source: C. Closed-loop Goal-conditioned Hair… · p. 6, §VIII · p. 8

C. Closed-loop Goal-conditioned Hair Styling

VIII. CONCLUSION

  • The paper introduces a model-based hair-care system and a 3D hair-combing dynamics model, supported by a strand-level, contact-rich simulator.The system is evaluated for visual goal-conditioned styling in simulation and the real world.Source: §VIII · pp. 7–8

IX. APPENDIX

The appendix points to additional implementation details and results, and describes a gravity-scheduling mechanism for stabilizing simulated hair.

A. Table of Contents

  • The appendix directs readers to further details on simulation, dynamics learning, closed-loop styling, additional results, failure modes, and system time costs.The listed material includes seen-versus-unseen hairstyle comparisons and further qualitative styling results.Source: A. Table of Contents · p. 10

B. More Details on Simulation

  • The simulation details include downsampling hairstyles to 2K strands and preprocessing strand density while preserving curvature.The appendix also points to a table of key simulation parameters; the twist-constraint index gap is tuned for each hairstyle.Source: B. More Details on Simulation · p. 10, Table II · p. 10
  • Gravity is enabled only for strands with sufficient positional movement, while other strands remain stable to reduce gravity-induced sagging.The mechanism supports visually realistic, physically plausible simulation for synthetic data generation and closed-loop evaluation.Source: B. More Details on Simulation · p. 10

C. More Details on Dynamics Learning

The section details the hierarchical volumetric dynamics model and its baselines, the processing of synthetic hair data, and metrics designed to assess positional, orientation, and strand-level prediction errors.

1) Method Implementation

  • The method hierarchically encodes occupancy and orientation with motion-aligned contact grids, fuses state and action features for latent editing, quantizes them, and decodes predicted end states.It uses two encoding levels, codebooks at both levels, and residual transposed convolutions for up-sampling.Source: 1) Method Implementation · p. 10
  • PC-GNN uses 2,000 hair points and 10 neighbors across 10 message-passing steps, predicting positions and orientations with separate supervision.Earth Mover’s Distance is omitted as computationally infeasible at the required point-cloud resolution.Source: 1) Method Implementation · pp. 10–11
  • V-UNet uses a four-level encoder-decoder with attention-based motion fusion and skip connections, while V-FiLM conditions shared pre-trained state encoders through channel-wise affine transformations.Neither baseline changes the method’s loss functions; V-UNet is not pretrained, whereas V-FiLM shares the method’s pretrained encoders, codebooks, and decoders.Source: 1) Method Implementation · p. 11

2) Data Processing

  • Synthetic states are converted from 2K strands to oriented point clouds, then down-sampled to 50K points and voxelized at 64×64×128 for volumetric methods or reduced to 2K points for PC-GNN.Each voxel represents approximately 5 mm in real-world physical space.Source: 2) Data Processing · p. 11

3) Formal Definitions of the Evaluation Metrics

  • Evaluation converts volumetric outputs to 2K-point clouds and measures positional Chamfer distance, orientation error, and strand-level Chamfer distance.The metrics include percentile statistics to capture sparse but significant deformations; orientation comparisons use nearest neighbors within a 2 cm positional threshold, and strand distance accounts for piecewise-linear segments.Source: 3) Formal Definitions of the… · pp. 11–12

D. Additional Results on Dynamics Learning

The results assess qualitative local deformation, compare performance on seen and unseen hairstyles, and test the contribution of hierarchical latent structure. DYMO-Hair generalizes better than the other methods on unseen hairstyles, while the hierarchy improves reconstruction and dynamics learning.

1) Qualitative Evaluation

  • Among the compared methods, DYMO-Hair best captures local hair deformation near the motion region on unseen hairstyles.The qualitative comparison shows initial states, ground-truth end states, and predictions; key regions are highlighted.Source: 1) Qualitative Evaluation · p. 12

2) Comparison on Seen and Unseen Hairstyles

  • On 7 unseen hairstyles, DYMO-Hair outperforms all other methods, while V-UNet fits the 10-hair training set best but generalizes poorly.DYMO-Hair is second-best on seen hairstyles, with only a small gap to V-UNet.Source: 2) Comparison on Seen and Unseen… · p. 12
  • V-UNet’s training-set advantage is attributed to skip connections that preserve input details, while quantization in DYMO-Hair may lose detail despite its pretrained latent space.The authors note that some fine-grained detail loss may remain and suggest improved latent-space pretraining as future work.Source: 2) Comparison on Seen and Unseen… · p. 12

3) Ablation Study on Hierarchical Model Structure

  • Compared with the single-layer variant, the hierarchical structure improves reconstruction accuracy on seen and unseen hairstyles and improves dynamics learning performance.The ablation also reports that the hierarchy preserves finer geometric details and supports a richer latent space.Source: 3) Ablation Study on Hierarchical… · p. 12

E. More Details on Closed-loop Hair Styling

The section describes perception and MPPI planning details, additional simulated styling results on unseen hairstyles, and limitations of the hair simulator. It reports that gravity and contact-model artifacts constrain some simulated outcomes.

E. More Details on Closed-loop Hair Styling

  • Real-world perception reconstructs 3D hair from 56 RGB-D viewpoints, while the rule-based system uses a front-view image and a 2D orientation map.Simulation comparisons instead provide oracle strand information to the systems.Source: E. More Details on Closed-loop Hair… · p. 12
  • Model-based planners sample 48 candidate 3D tool motions and use MPPI over a two-step horizon, executing the first step to minimize strand-level distance to the goal.The rule-based planner is reimplemented from the prior work’s high-level methodology because its original code is unavailable.Source: E. More Details on Closed-loop Hair… · pp. 12–13

F. Additional Results on Closed-loop Hair Styling

  • Additional simulation examples show DYMO-Hair styling diverse unseen hairstyles from messy initial conditions, though some cases leave a slight goal gap within the 15-step budget.The figure presents two cases per hairstyle, from goal and initial state through planning steps to end state.Source: F. Additional Results on Closed-loop… · p. 13

1) Simulation

  • The PBD simulator provides a lightweight method for strand-level, contact-rich hair-combing simulation used for data generation and closed-loop experiments, but can produce artifacts.The authors characterize it as visually realistic and physically plausible overall.Source: 1) Simulation · p. 13
  • Disabling gravity to prevent sagging while preserving initial curvature with the twist constraint can leave some long, complex strands appearing suspended.The authors note this artifact in a simulated example.Source: 1) Simulation · pp. 13–14
  • Particle-based collision handling may inflate hair when initial inter-strand gaps are smaller than particle size, limiting the simulator’s use beyond dynamic combing.The authors identify realistic combing dynamics, rather than exact preservation of the initial hairstyle, as the simulator’s primary focus.Source: 1) Simulation · p. 14

2) Dynamics Modeling

The authors identify remaining limits in dynamics prediction and closed-loop styling, including detail loss, a scalar planning cost, gravity, and tool contact. They also suggest that more diverse data and greater model capacity may improve generalization and robustness.

2) Dynamics Modeling

  • Voxel-level supervision can leave gaps or discontinuities in predicted unseen-hair end states because it lacks geometric-continuity regularization.The authors suggest strand-structure priors or geometric regularization as possible remedies.Source: 2) Dynamics Modeling · p. 14
  • The hierarchical architecture improves detail preservation, but some fine-grained geometry may still be lost.Future work could improve latent-space pretraining to balance generalizability and geometric detail fidelity.Source: 2) Dynamics Modeling · p. 14
  • More diverse hairstyles and dynamics data, together with greater model capacity, may improve representation, generalizability, and robustness across a wider range of hairstyles..Source: 2) Dynamics Modeling · p. 14

3) Hair Styling System

  • A single strand-level distance can inadequately represent hair-state differences, sometimes leaving styling incomplete within the step budget.Combining complementary costs may help, but balancing their weights could add complexity.Source: 3) Hair Styling System · p. 14
  • In 3D styling, gravity can make hair fall back when motions create insufficient occlusion or friction; closed-loop replanning and retries help but are not optimal.Future options include modeling gravity or sampling actions with head geometry in mind.Source: 3) Hair Styling System · p. 14
  • The deformable fingertip can remain too stiff, requiring millimeter-level calibration and risking excessive force or strands slipping through tool-head gaps.Softer compliant tools may reduce calibration sensitivity but could alter contact dynamics and complicate modeling.Source: 3) Hair Styling System · pp. 14–15

H. System Time Cost Analysis

A five-step real-world manipulation experiment takes approximately 60 minutes, and the analysis identifies view capture, state estimation, and planning as targets for improving system efficiency.

H. System Time Cost Analysis

  • Approximately 60 minutes are required for a five-step real-world manipulation experiment, including five observation-planning-execution cycles and a final observation for evaluation.Each cycle includes multi-view observation, action planning, and execution.Source: H. System Time Cost Analysis · p. 15
  • 56 views are captured in about 3 minutes, followed by around 4 minutes of state estimation for each multi-view observation.The robot takes 3 seconds to move and capture each view.Source: H. System Time Cost Analysis · p. 15
  • Each action step takes approximately 2.5 minutes for planning and 30 seconds for execution.Source: H. System Time Cost Analysis · p. 15
  • Reducing the 56 dense views could improve efficiency, but requires balancing view count against state-estimation accuracy.The current dense capture is used to ensure accurate state estimation.Source: H. System Time Cost Analysis · pp. 15–16
  • Parallelization and GPU acceleration could shorten state estimation and planning, where large-batch sampling and strand-level distance computation remain time-consuming.State estimation is already much faster than full strand reconstruction, which typically takes tens of minutes.Source: H. System Time Cost Analysis · p. 16
Loading 2510.06199v2…