Geometry-Aware Reinforcement Learning for Close-Fitting Wearable Insertion

Takuma Tsukakoshi; Elle Miller; Tamon Miyake; Namiko Saito; Yushi Wang; Shigeki Sugano; Tetsuya Ogata; Sethu Vijayakumar

cs.ROOct 23, 2026

Geometry-Aware Reinforcement Learning for Close-Fitting Wearable Insertion

Authors: Takuma Tsukakoshi, Elle Miller, Tamon Miyake, Namiko Saito, Yushi Wang, Shigeki Sugano, Tetsuya Ogata, Sethu Vijayakumar

Organizations: Waseda University · University of Edinburgh · Microsoft Research Asia

Abstract

Robot-assisted dressing is particularly challenging due to narrow openings and articulated body parts. Even small errors in the relative pose between the body part and garment opening can cause snagging, missed insertion, or incomplete dressing. To address these challenges, we formulate dressing as a geometry-aware insertion problem that represents task-relevant spatial relationships between the body part and object opening, rather than relying on the full deformable-object state. Based on this formulation, we propose a Reinforcement Learning (RL) framework combining a staged curriculum, which progressively guides the policy from coarse alignment to insertion and final placement, with a finger-wise insertion soft gate that provides continuous feedback on individual finger insertion progress and helps suppress premature insertion. We evaluate the framework on rigid and deformable bracelet dressing tasks through component ablations, systematic variations in the initial hand pose, and closed-loop execution of the learned policies on a real robotic system. On the post-training deformable evaluation, the full method reaches 84.7 +- 1.2% success. The insertion gate accounts for the largest gain, and adding the curriculum on top of that gate reaches the top score. On a mannequin hand the same full policy succeeds on 15/20 real-robot trials.

Figures & tables

Fig. 1: Real-world closed-loop dressing. Camera keypoints of the hand and opening guide reaching, insertion, and placement on a mannequin.

Fig. 2: Method overview. The policy uses robot proprioception, user hand keypoints, and wearable-object keypoints to encode relative hand-object geometry and generate dressing actions. The primary evaluation is the Deformable Bracelet task; a Rigid Bracelet diagnostic isolates insertion when the opening cannot collapse.

Fig. 3: Curriculum and the finger-wise soft gate. (a) Little-side, thumb-side, and wrist distances define the staged reward from coarse alignment to wrist insertion. (b) The soft gate uses the top, center, and bottom opening keypoints and saturates as more fingertips enter the opening height band.

Type	Symbol	Description	Dim.
$\mathbf{o}^{\mathrm{prop}}_{t}$	$\bar{\mathbf{q}}_{t}$	normalized arm and torso joint positions	$17$
	$\bar{\dot{\mathbf{q}}}_{t}$	normalized arm and torso joint velocities	$17$
	$\mathbf{e}^{q}_{t}$	arm and torso joint position errors	$17$
	$\mathbf{x}^{R}_{\mathrm{EE},t}$	right end-effector position	$3$
	$\mathbf{r}^{R}_{\mathrm{EE},t}$	right end-effector orientation	$4$
	$\mathbf{x}^{L}_{\mathrm{EE},t}$	left end-effector position	$3$

TABLE I: Observation space. The policy input

\mathbf{o}_{t}

consists of robot proprioceptive and task-relevant geometric features.

$r_{k}$	$D_{k}$	$\sigma_{k}$	$\lambda_{k}$	$a_{k}$
$r_{\mathrm{ee\_little}}$	$\\|\mathbf{x}_{\mathrm{left\_ee}}-\mathbf{x}_{\mathrm{little}}\\|$	$0.05$	$10.0$	$c_{\mathrm{ee}}\,c_{\mathrm{little}}$
$r_{\mathrm{ee\_thumb}}$	$\\|\mathbf{x}_{\mathrm{right\_ee}}-\mathbf{x}_{\mathrm{thumb}}\\|$	$0.10$	$20.0$	$c_{\mathrm{ee}}\,c_{\mathrm{thumb}}$
$r_{\mathrm{wrist\_opening}}$	$\\|\mathbf{x}_{\mathrm{wrist}}-\mathbf{x}_{\mathrm{opening}}\\|$	$0.14$	$200.0$	$c_{\mathrm{ee}}\,g_{\mathrm{fingers}}$

TABLE II: Reward components and activation gates for wearable-object dressing.

Fig. 4: Real-world setup. A head-mounted egocentric camera estimates hand keypoints and the top opening marker, while an external camera observes the bottom marker. Both markers are tracked with SAM2. All views are transformed into the robot base frame.

Fig. 5: Mean evaluation success rate during training for the rigid and deformable bracelet dressing tasks. Solid lines are the three-seed mean; shaded regions are the min–max range. These curves are not the post-training $50$ -episode scores in Table III . Full (soft gate + curriculum) is highest on both panels. The insertion gate accounts for the large gap; the deformable ceiling is lower than the rigid diagnostic because the opening can collapse.

Fig. 6: Success criterion and failure modes. Left: success requires all five finger bases to cross the opening plane within the ellipse and the wrist-to-opening distance to be $<1\,\mathrm{cm}$ . Right: typical failures: snagging, partial insertion, overshooting, and excessive deformation.

Deformable Bracelet
Method	Success (%)	Fingers	Deflection ( ∘ )
Full (Soft Gate + Curriculum)	$\boldsymbol{84.7\pm 1.2}$	$4.85\pm 0.11$	$0.74\pm 0.18$
Hard Gate + Curriculum	$76.7\pm 5.0$	$4.78\pm 0.05$	$0.74\pm 0.10$
Soft Gate + No Curriculum	$71.3\pm 9.9$	$4.77\pm 0.08$	$0.57\pm 0.15$
Curriculum + No Gate	$60.7\pm 7.0$	$3.62\pm 0.38$	$2.30\pm 0.35$
No Gate + No Curriculum	$48.0\pm 9.2$	$2.98\pm 0.24$	$1.63\pm 1.10$

TABLE III: Post-training ablation evaluation of each seed’s best checkpoint (highest mean evaluation return). Mean

\pm

std over three seeds;

50

episodes per seed.

Fig. 7: Inserted-finger count at the final step. Deformable Bracelet post-training evaluation; three-seed mean over 50 episodes.

Condition	Range	Success	Fingers	Deflection ( ∘ )
Pos.	$\pm 2.5\,\mathrm{cm}$	$44/50$	$4.92\pm 0.40$	$0.44\pm 0.29$
	$\pm 5.0\,\mathrm{cm}$	$46/50$	$4.98\pm 0.14$	$0.47\pm 0.31$
	$\pm 7.5\,\mathrm{cm}$	$40/50$	$4.92\pm 0.44$	$0.76\pm 0.51$
Yaw/Pitch	$\pm 5^{\circ}$	$49/50$	$4.98\pm 0.14$	$0.34\pm 0.47$
	$\pm 10^{\circ}$	$46/50$	$4.76\pm 0.74$	$0.60\pm 1.24$
	$\pm 15^{\circ}$	$39/50$	$4.82\pm 0.83$	$1.06\pm 1.20$

TABLE IV: Pose robustness on the Deformable Bracelet task. Success is reported as count

/50

episodes.

Method	Success
Full (Soft Gate + Curriculum)	$15/20$ ( $75\%$ )
Hard Gate + Curriculum	$12/20$ ( $60\%$ )
Soft Gate + No Curriculum	$10/20$ ( $50\%$ )
Curriculum + No Gate	$9/20$ ( $45\%$ )
No Gate + No Curriculum	$6/20$ ( $30\%$ )

TABLE V: Real-robot success across

20

mannequin poses, with one closed-loop trial per pose.

Robotics

Robotics studies the design and control of machines that perceive, move, and act in physical environments.

See all 747 papers

Dressing in Motion: A Human Motion-Aware Diffusion Policy for Robot-Assisted Dressing
Sep 4, 2026Haoxiang Sun, Fangyuan Wang, Songhao Huang +4Assistive Robotics Dynamic Robotic Manipulation
Department of Mechanical Engineering, The Hong Kong Polytechnic University, Hong Kong SAR, China. · School of Nursing, The Hong Kong Polytechnic University, Hong Kong SAR, China. · School of Physics, Engineering and Technology, University of York, UK. +1
InsertAnything: Generalizable Contact-Rich Precision Insertion from Simulation to Reality
Sep 21, 2026Zhenghua Ma, Xinpan Meng, Zeyu Liu +4RL for Robotics Contact-Rich Robotic Manipulation
School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China, and the State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · PaXini AI Technology (Beijing) Co., Ltd., Beijing 100080, China
Instrumentation for Imitation Learning: Enhancing Training Datasets for Clothes Hanger Insertion
May 22, 2026Remko Proesmans, Thomas Lips, Francis wyffelsRobot Skill Learning Robotic Insertion
AI and Robotics Lab (IDLab-AIRO), Ghent University—imec, Ghent, Belgium

Sep 4, 2026J/K move · Enter open · S save

Dressing in Motion: A Human Motion-Aware Diffusion Policy for Robot-Assisted Dressing

Haoxiang Sun, Fangyuan Wang, Songhao Huang, Justina Y. W. Liu+3

Department of Mechanical Engineering, The Hong Kong Polytechnic University, Hong Kong SAR, China. · School of Nursing, The Hong Kong Polytechnic University, Hong Kong SAR, China. · School of Physics, Engineering and Technology, University of York, UK.+1

Assistive Robotics Dynamic Robotic Manipulation Diffusion Policy Robot Imitation Learning

Open paper AskPDF arXiv

Robotic dressing assistance is a promising solution for supporting older adults with physical impairments in daily living. However, dressing under human motion remains challenging, as complex garment--human contact and occlusions make it difficult to generate actions aligned with arm movements. In this letter, we propose a visuomotor policy that learns dressing skills from static expert demonstrations and generalizes to dynamic user-motion scenarios. A diffusion policy tailored to garment--human interaction geometry learns from partially observed point clouds with varied arm postures. We then introduce an object-centric representation based on PDE diffusion to capture the axial distribution of the arm. By sampling motion-relevant regions and registering them across consecutive observations, the proposed method approximates arm motion and reactively adapts the executed trajectory. We evaluate our method in simulation and a real-world human study involving nine participants, three garment types, and six arm-motion patterns. Results show that our method outperforms baselines in dressing progress, freedom of movement, and user comfort. The project website is https://anonymous.4open.science/w/dressing-in-motion.

Abstract

Robotics

Explore similar work

Dressing in Motion: A Human Motion-Aware Diffusion Policy for Robot-Assisted Dressing

InsertAnything: Generalizable Contact-Rich Precision Insertion from Simulation to Reality

Instrumentation for Imitation Learning: Enhancing Training Datasets for Clothes Hanger Insertion