Assistive robots have the potential to provide tremendous support for disabled and elderly people in their daily dressing activities. This thesis presents a series of perception and manipulation algorithms for robot-assisted dressing, including: garment perception and grasping prior to robot-assisted dressing, real-time user posture tracking during robot-assisted dressing for (simulated) impaired users with limited upper-body movement capability, and finally a pipeline for robot-assisted dressing for (simulated) paralyzed users who have lost the ability to move their limbs.

First, the thesis explores learning suitable grasping points on a garment prior to robot-assisted dressing. Robots should be endowed with the ability to autonomously recognize the garment state, grasp and hand the garment to the user and subsequently complete the dressing process. This is addressed by introducing a supervised deep neural network to locate grasping points. To reduce the amount of real data required, which is costly to collect, the power of simulation is leveraged to produce large amounts of labeled data.

Unexpected user movements should be taken into account during dressing when planning robot dressing trajectories. Tracking such user movements with vision sensors is challenging due to severe visual occlusions created by the robot and clothes. A probabilistic real-time tracking method is proposed using Bayesian networks in latent spaces, which fuses multi-modal sensor information. The latent spaces are created before dressing by modeling the user movements, taking the user's movement limitations and preferences into account. The tracking method is then combined with hierarchical multi-task control to minimize the force between the user and the robot. The proposed method enables the Baxter robot to provide personalized dressing assistance for users with (simulated) upper-body impairments.

Finally, a pipeline for dressing (simulated) paralyzed patients using a mobile dual-armed robot is presented. The robot grasps a hospital gown naturally hung on a rail, and moves around the bed to finish the upper-body dressing of a hospital training manikin. To further improve simulations for garment grasping, this thesis proposes to update more realistic physical properties values for the simulated garment. This is achieved by measuring physical similarity in the latent space using contrastive loss, which maps physically similar examples to nearby points.