Master-slave dual-arm learning from demonstration assembly based on modified triple reversible dynamic motion primitives

For the peg-in-hole assembly tasks with small clearance, the traditional robot trajectory generation methods rely heavily on expert knowledge, which is complex and costly. In contrast, learning from demonstration method does not require expert knowledge, and can enable the robots to quickly learn and encode assembly trajectories. However, due to the rich contact states, easy jamming and collaborative constraints during the assembly process, dual arm learning from demonstration assembly remains a challenging task. For the reason, this paper first proposes a new dynamic motion primitives method that has both global asymptotic stability and reversibility, which can switch the forward and reverse directions of the trajectory generation process at any time to alleviate the jamming problem. Then, a uniform and concise robot pose synchronization description approach based on triple position reversible dynamic motion primitives is given. On this basis, through utilizing triple reversible dynamic motion primitives for the dual arm collaborative trajectory learning, and introducing slave-arm force coupling terms to modify them for trajectory compliance, a master-slave dual-arm learning from demonstration assembly algorithm is provided. Finally, based on two UR5 robots, a series of assembly experiments with three different shapes of pegs and holes are carried out, which confirm the effectiveness of the proposed method.