Robot-Assisted Dressing System

• Robot-assisted dressing systems are advanced robotic solutions that enable safe and adaptive garment donning through human–robot interaction.
• They utilize reinforcement learning, sensor-based pose estimation, and haptic model predictive control to manage non-rigid materials and ensure user safety.
• Real-world implementations integrate multi-modal sensor fusion, bimanual/soft robotic architectures, and formal safety frameworks to personalize assistance.

A robot-assisted dressing system is an embodiment of physical human–robot interaction that enables a robot to aid individuals—particularly those with limited mobility—in donning garments. Such systems address challenges inherent to manipulation of non-rigid materials, safe physical contact with humans, and accommodation of diverse user abilities and behavioral responses. The research landscape for robot-assisted dressing encompasses reinforcement learning for modeling human behaviors, sensor-based pose estimation, force-aware control, multi-modal and personalized policies, bimanual or soft robotic architectures, and formal safety frameworks. The following sections survey key methodologies, evaluation protocols, control architectures, and future research trajectories as evidenced in leading work on the topic.

1. Learning Human Behavior and Policy Optimization

A central paradigm in robot-assisted dressing is the modeling of human behavior using reinforcement learning (RL). Rather than replicating recorded human motion or relying on handcrafted rules, RL enables the autonomous discovery of control policies that are capable of placing a human limb into a garment under various forms of robot assistance (Clegg et al., 2017).

The human dressing controller is formulated as a partially observable Markov Decision Process (POMDP), where the policy observation vector—163-dimensional—incorporates proprioceptive cues, garment feature locations (e.g., sleeve centroid and hand-sleeve displacement), aggregated haptic feedback, surface contact sign, and a task vector. Critically, this design restricts access to information so that the learned policies act upon cues realistically accessible by humans.

Policy optimization leverages Trust Region Policy Optimization (TRPO), ensuring stable updates even when rewards are sparse and the action space is high-dimensional and continuous. Training typically demands 1000–2000 TRPO iterations. The learned policies aim to determine what is achievable by human motion rather than merely what is observed in practice, supporting evaluations across different robot assistance strategies—fixed sleeve, front linear motion, side linear motion.

The reward function combines progress (quantified by insertion depth in the sleeve using containment and intersection checks), deformation penalty (based on triangle mesh strain in the simulated gown), geodesic reward (on the cloth surface), and posture maintenance:

$$r(s) = w_1 \cdot r_p(s) + w_2 \cdot r_d(s) + w_3 \cdot r_g(s) + r_u(s)$$

where weights are empirically calibrated (e.g., $w_1=5, w_2=6, w_3=2$).

2. Sensor-Based Pose Estimation and Real-Time Adaptation

Accurate estimation of the user's pose during dressing is made difficult by visual occlusions and unpredictable user motion. Capacitive proximity sensing methods have been advanced to overcome these limitations (Erickson et al., 2017, Erickson et al., 2019). Single-axis and multidimensional capacitive sensors provide direct, low-latency measurements of proximity, robust to occlusion and effective even through clothing.

A canonical implementation employs a copper-foil coated electrode sampling capacitance at $\geq 200$ Hz. The parallel plate model yields a closed-form distance estimate:

$d(\Delta C) = \frac{\alpha}{\Delta C + \beta}$

with calibration parameters (e.g., $\alpha \approx 84.38, \beta \approx 4.681$). Feedback control structures—PD controllers—maintain desired offsets (e.g., 5 cm) between the robot end effector and the limb, dynamically compensating for errors in initial pose estimate and tracking vertical or lateral human movement.

A neural approach provides further generality: a feed-forward network processes a temporal window of multidimensional capacitance measurements to estimate limb translation and orientation, enabling real-time PD control for both dressing and bathing tasks (Erickson et al., 2019). This multidimensional capacitive sensing is particularly effective for physical human–robot interaction tasks due to its robustness to environmental factors and low computational overhead.

3. Force Prediction and Haptic Model Predictive Control

Safety and comfort hinge upon the robot’s ability to predict and regulate forces imparted to the user during dressing. Deep haptic model predictive control frameworks have been introduced in which recurrent networks (e.g., pairs of three-layer LSTMs with 50 cells per layer) jointly predict force maps on the human body from sequences of haptic/kinematic observations and candidate action sequences (Erickson et al., 2017). The estimator network outputs taxel-specific contact forces, while the predictor forecasts future measurements given proposed actions.

This architecture is embedded within a model predictive control (MPC) loop, minimizing an objective function penalizing predicted high forces (via squared $L_1$ norm), unwanted rotations, and rewarding forward movement:

$$J(x_{1:t}, a_{t+1:t+H_p}) = w_1 \|F(x_{1:t}, G(x_{1:t}, a_{t+1:t+H_p}))\|_1^2 - w_2 \sum_{j=t+1}^{t+H_p} (\bar{d} \cdot a_{j,v}) + w_3 \sum_{j=t+1}^{t+H_p} |a_{j,\psi}|$$

Self-supervised simulation (e.g., 10,800 trials) provides extensive coverage, mitigating real-world risk and supporting model evaluation and personalization. Longer prediction horizons (0.2 s) translate to high dressing completion rates (e.g., $97.5-98.75\%$) and lower force application than short-horizon methods, which are prone to garment snagging at joints.

4. Control Architectures and Strategy Adaptation

Recent advances focus on hybrid and adaptive control strategies combining multiple feedback modalities, user interaction, and runtime verification:

• Hybrid Low-level Controls: These combine continuous force monitoring (e.g., thresholds $F > 35\,\mathrm{N}$ trigger pausing and recovery) with interaction-based recovery, via natural language chatbots for soliciting user intervention or safe abort (Rafiq et al., 12 May 2025).
• Runtime Verification: Safety is encoded through a parametric discrete-time Markov chain (pDTMC) model of the dressing process, with transitions dynamically updated via Bayesian inference. Safety constraints are expressed in probabilistic computation tree logic (PCTL) and verified symbolically (e.g., $P_{\leq 0.1}[F s = 8]$ restricts abort probability); closed-form expressions are precomputed for real-time evaluation (Rafiq et al., 22 Apr 2025).
• Personalization and User Agency: Systems such as GRACE employ functional range of motion (fROM) embeddings, predicting feasible limb configurations contingent on clinical assessment scores via autoencoder-based latent representations. Feasibility predictions guide planning so that dressing assistance matches user capabilities and preserves agency (Liu et al., 29 Jan 2025).

5. Bimanual, Soft Robotic, and Specialized Dressing Systems

Single-arm robotic methods are commonly effective for loose-fitting garments but fail for tighter clothing due to armscye constraints. Bimanual schemes deploy paired robots—one manipulating the garment, the other guiding the limb or supporting the arm—enabling adaptive tensioning, pose tracking, and improved dressing efficacy, especially for narrow or tight sleeves (Zhu et al., 2023, Zhao et al., 17 Aug 2025). Spherical or dressing-relative coordinate systems, learned via Gaussian Mixture Model (GMM) and Gaussian Mixture Regression (GMR), encode posture-adaptive bimanual trajectories.

Soft robotic systems represent an alternative, using deployable fabric sheaths actuated by internal "subvines" (pressurized soft actuators). The Self-Wearing Adaptive Garment (SWAG) unfurls around the limb (driven by tip growth and a distributed, nonlinear friction pulley mechanism), eliminating rigid control, minimizing skin-garment friction, and achieving rapid, posture-conforming donning (e.g., $14$ s for a sleeve) (Kim et al., 9 Jul 2025).

Systems increasingly address difficult scenarios—close-fitting socks (hierarchical LSTM fusion of vision, proprioception, and tactile input with semantic attention and depth estimation) (Tsukakoshi et al., 6 May 2025), as well as the pre-dressing step (garment unfolding) via imitation-learned manipulation primitives (e.g., dynamic “fling,” “shake,” and controlled “twist,” parameterized by constrained dynamic movement primitives and state classification) (Blanco-Mulero et al., 24 Jul 2025).

6. Robustness to Real-World Constraints and Multi-Modal Feedback

Virtually all systems face the sim-to-real gap and occlusion challenges. Strategies include:

• Multi-modal Policy Augmentation: Vision-based policies are fine-tuned using force feedback (via FiLM layers) to produce force-modulated visual policies (FMVP) (Hao et al., 16 Sep 2025). This structure enhances adaptation to dynamic limb movement and occlusion beyond simulation capabilities.
• Constrained Action Selection: Safety constraints are enforced via force dynamics models trained on real-world data. Action proposals from vision-based policies are filtered by force predictions, ensuring candidate selections never exceed safe force thresholds (Sun et al., 2023).
• Comprehensive Personalization: User-specific limb mobility (personalized fROM) is incorporated directly into planning using functional assessment embeddings (Liu et al., 29 Jan 2025).
• Wearable Interfaces: Head-worn (HAT) and teleoperation systems, as well as shared and passive control integrations, facilitate user-driven and user-informed robot actions (Padmanabha et al., 7 Feb 2025).

These approaches underpin robust policy generalization to unseen garments and user poses, successful sim-to-real transfer, and enhanced user safety and comfort across diverse experimental setups.

7. Evaluation, Metrics, and Future Research Directions

Protocols for evaluation span simulated physics environments (e.g., DART, PhysX, NVIDIA FleX/SoftGym), manikin and physical human trials, and are measured via quantitative (arm/garment progress ratio, applied force metrics, cloth deformation thresholds, success rate) and qualitative (Likert-scale user satisfaction, agency, discomfort) criteria. Noteworthy metrics include the arm/upper-arm dressed ratio, force violation rates, and "dressing effectiveness indicators" (e.g., arm coverage projections), as well as statistical significance measures (e.g., Wilcoxon tests for user paper outcomes).

Future developments are expected to extend:

• Integration of advanced sensory modalities (high-resolution tactile, multi-view vision, real-time depth sensing).
• Further exploration of recurrent, memory-augmented RL architectures to better handle sequential and dynamic dressing scenarios.
• More aggressive domain randomization and policy distillation for robust transfer across garment types and user postures (Wang et al., 2023).
• Safety guarantees via formal runtime verification and adaptive control logic (Rafiq et al., 22 Apr 2025).
• Stronger personalization to user abilities and preferences via fROM prediction, real-time feedback, and active learning (Liu et al., 29 Jan 2025).
• Expansion from sleeve dressing scenarios to more complex tasks (e.g., bilateral dressing, over-the-head shirts, zipped garments) and broader application domains (medical, industrial, emergency protective gear) (Kim et al., 9 Jul 2025).

These directions will be characterized by increasingly sophisticated multi-modal sensor fusion, physically compliant actuation, formal safety assurances, and adaptive, context-sensitive behavior—defining the next generation of robot-assisted dressing systems.