RD-GMM: A GMM-Based Framework for Robotic Dressing with a Single Fixed Robotic Arm Applied to T-Shirt Wearing

• Maria Fernanda Paulino Gomes UNICAMP
• César Bastos da Silva UNICAMP
• Ervin Alain Huayhua UNICAMP
• Felipe Augusto Mota UNICAMP
• Leonardo Rocha Olivi UFJF
• Eric Rohmer UNICAMP

Resumo

Robotic assistance for dressing remains one of the most challenging tasks in assistive robotics due to the complex dynamics of deformable objects and the need for precise coordination around human bodies. Many recent works combine vision-based keypoint detection with learning-based motion generation, but they often rely on strong assumptions or large training datasets. Motivated by the need for reproducible dressing frameworks, this work proposes a modular framework for autonomously dressing t-shirts using a robotic manipulator, Robotic Dressing Gaussian Mixture Modeling (RD-GMM). The system integrates RGB-D perception for body keypoint detection, motion generation based on Gaussian Mixture Models (GMM) and Gaussian Mixture Regression (GMR), which learn trajectories from a limited number of kinesthetic demonstrations, all managed by a state machine. The framework is experimentally validated on a physical mannequin using a Kinova Gen3 robot, demonstrating high system performance in dressing a t-shirt task with the inclusion of GMM. Despite the challenges of handling deformable objects and the assumption of a static human posture, the results show the feasibility of dressing a human with a single manipulator without the need for regrasping.

Palavras-chave: Training, Adaptation models, Visualization, Robot kinematics, System performance, Clothing, Real-time systems, Trajectory, Robots, Gaussian mixture model, Dressing, Gaussian Mixture Models, Cloth Manipulation, Perception, Human Robotic Interaction