A Novel Collaborative Imitation Learning Framework for Dual-Arm Flipping Tasks

Imitation learning encodes skills through human-friendly demonstration, enabling the acquisition of complex skills. During the teleoperation-based demonstration process for dual-arm tasks, operators typically need to simultaneously control two master devices, which will greatly impact the quality of the operation. Furthermore, the dual-arm flipping task includes inherent collaborative constraints. When traditional models are applied for separate learning, the success rate of generalization tends to be low, and redundant degrees of freedom can hinder efficiency. In this paper, we propose a teleoperation-based assisted flipping strategy. The operator only needs to intuitively control the motion of one manipulator, while another manipulator is able to subsequently adjust its pose and radial contact force to achieve the stable flipping of the object. Considering the pose constraints of the two manipulators, we design a new generalized position describing the task progression and use it to drive the dynamical system control outputs for both manipulators. Ultimately, we conduct experiments to verify the effectiveness and robustness of our collaborative imitation learning framework for the flipping task, resulting in an improved success rate of task generalization.