Neuroadaptive control for safe robots in human environments: A case study

Safety is an important consideration during physical Human-Robot Interaction (pHRI). Recently the community has tested numerous new safety features for robots, including accurate joint torque sensing, gravity compensation, reduced robot mass, and joint torque limits. Although these methods have reduced the risk of high energy collisions, they rely on reduced speed or accuracy of robot manipulators. Indeed, because lightweight robots are capable of higher velocities, knowledge of dynamical models is required for precise control. However, feedforward compensation is difficult to implement on lightweight robots with flexible and nonlinear joints, links, cables, and so on. Furthermore, unknown objects picked up by the robot will significantly alter the dynamics, leading to deterioration in performance unless high controller gains are used. This paper presents an online learning controller with convergence guarantees, that is able to learn the robot dynamics on the fly and provide feed-forward compensation. The resulting joint torques are significantly lower than conventional independent joint control efforts, thus improving the safety of the robot. Experiments on a PR2 robot arm are conducted to validate the effectiveness of the neuroadaptive controller to reduce control torques during high speed free-motion, lifting unknown objects, and collisions with the environment.