A Telerobotic Shared Control Architecture for Learning and Generalizing Skills in Unstructured Environments

Direct teleoperation of robots in unstructured environments by non-experts often leads to low efficiency and increased risk. To this end, this paper proposes a shared control architecture where the robot can generalize demonstrations based on variable environments (start, obstacles, and goal positions) and infer user intention online to assist tasks safely and efficiently. First, the complex task is decomposed into unit actions, where the cubic-quintic-cubic Bezier curve is utilized to resolve the limitation of the classical dynamic movement primitives (DMP) algorithm in generalizing via-points. Gaussian process regression (GPR) was implemented to generalize expert demonstrations to different environments, ensuring full generalizable trajectory in the subtask space. GPR further extrapolate simple demonstrations from a subspace to the entire task space, avoiding the demand for numerous original demonstrations. Then, the online evolution of DMP is optimized: 1) an adaptive temporal scaling system is developed to synchronize evolution with human operations; 2) an intention prediction and expected evolution selection method based on operational input is proposed, achieving human-led operation guidance. Experiments validate the effectiveness of the generalization, obstacle avoidance, and operation assistance. Tests with non-experts reported the designed architecture enhances safety and efficiency, and reduces collisions. Note to Practitioners—This paper aims to improve the safety and efficiency of teleoperated robots in unstructured environments. To tackle complex missions with defined subtasks, this article adds auxiliary operation functions to the traditional direct control mode. The novel system can offline plan obstacle avoidance trajectories in unstructured environments and then guide human operations online. It can provide expected paths in advance, judge human intentions, and adapt to user execution speed. In practical applications, GPR can be trained from little number of demonstrations performed in a simple space. The guidance trajectories in complex task can be obtained by configuring multiple subspace units. With the proposed tactile feedback method, human operations can be effectively assisted.