Reinforcement Learning based Optimization for Cobot's Path Generation in Collaborative Tasks

Task-Parameterized Learning from Demonstrations (TP-LfD) is an effective approach for collaborative robot (cobot). It aims at generating a path of a cobot moving in a dynamic collaborative task (e.g., a pick-and-place task) adaptively with respect to knowledge learnt from demonstrated tasks. That is, the learnt knowledge from demonstrated tasks are considered task parameters, which are critical input for TP-LfD to generate a movement path of a cobot for a new dynamic task. To further enhance the adaptability of TP-LfD, in this paper, an improved TP-LfD ( $i$ TP-LfD) approach over other developed TP-LfD approaches is presented. One of the major contributions in $i$ TP-LfD is that a reinforcement learning based optimization algorithm is designed to eliminate irrelevant task parameters identified in demonstrations, which boosts the overall computational performance of cobot's path generation. In the end, case studies were used to validate and highlight the adaptability and robustness of the approach.