Imitation Learning and Teleoperation Shared Control With Unit Tangent Fuzzy Movement Primitives

Imitation learning is pivotal for robot skill acquisition and holds potential for teleoperation shared control applications. This article introduces a novel imitation learning method and shared control framework using unit tangent fuzzy movement primitives (UTFMP). Using unsupervised demonstration segmentation results obtained by UTFMP as membership function references, and employing type-2 fuzzy models to manage uncertainties in segmented subtrajectories across multiple demonstrations, we propose an imitation learning method. This method facilitates trajectory learning based on either time or distance, supports real-time trajectory modulation using inputs, such as time, position, or velocity, adeptly adapts to dynamic scenarios, and possesses the capability to seamlessly blends diverse trajectories. Then, a shared control framework with segmented subtrajectories and modulated trajectories is established to extend UTFMP's application. This framework includes intent recognition through fuzzy fusion and a Markov transition matrix to identify intent of operator, as well as shared arbitration to integrate operations from both operator and robot agent. Simulation experiments validate UTFMP's effectiveness for demonstration segmentation, imitation learning and modulation. Teleoperation experiments demonstrate increased trajectory smoothness and accuracy by at least 24.38% and 9.48%, respectively, a minimum 16.56% reduction in control time compared to conventional velocity control using the shared control proposed.