Learning garment manipulation policies toward robot-assisted dressing

• DOI: 10.1126/scirobotics.abm6010
• Publication Date: 2022-04-06T17:56:02Z
• Authors: Fan Zhang; Yiannis Demiris
• Keywords: Leverage (statistics)

Assistive robots have the potential to support people with disabilities in a variety of activities daily living, such as dressing. People who completely lost their upper limb movement functionality may benefit from robot-assisted dressing, which involves complex deformable garment manipulation. Here, we report dressing pipeline intended for these and experimentally validate it on medical training manikin. The is composed robot grasping hospital gown hung rail, fully unfolding gown, navigating around bed, lifting up user's arms sequence finally dress user. To automate this pipeline, address two fundamental challenges: first, learning manipulation policies bring an uncertain state into configuration that facilitates robust dressing; second, transferring object learned simulation real world leverage cost-effective data generation. We tackle first challenge by proposing active pre-grasp approach learns isolate area before grasping. combines prehensile nonprehensile actions thus alleviates grasping-only behavioral uncertainties. For second challenge, bridge sim-to-real gap policy transfer approximating simulator real-world physics. A contrastive neural network introduced compare pairs simulated observations, measure physical similarity, account parameters inaccuracies. proposed method enables dual-arm put back-opening gowns onto manikin success rate more than 90%.