Teaching Garment Manipulation to a Dual-Arm Robot Manipulator thought Reinforcement Learning and Human robot Collaboration

In this talk, we present a teaching-based system for garment manipulation with a dual arm robot manipulator. First of all we propose a reinforcement learning system to learn how to manipulate garments in a simulated environment. But as simulation environments still struggle to perfectly replicate the real-world features of fabrics and, due to the conditions of the working space on a factory can change under demand, we propose to adapt robotic arms to real-world conditions through direct human guidance enabling the transfer of learned knowledge from simulation to real-world operation. For this purpose, we propose a multi-critic reinforcement learning algorithm that can learn in a simulated environment and an adaptative hybrid impedance control to transfer this learning to the real world while protecting the clothes from damage. With this proposal we aim to bridge the gap between simulation and reality and enhance the applicability of simulation-trained models to complex real-world scenarios.