Learning Gravitation Compensation on a Simulated Passive Compliant Robot Limb for Multiple Targets and Load Cases

Compliance is an important feature for service robots but also for robots in industrial applications to increase human machine interaction safety. While active compliance is online adaptable and easy to control its dynamic response is determined by the sampling rate and the response rate of the controller. In contrast, passive compliance is inherent to the system. It responds naturally fast to any perturbation exerted on the robot independently of the controller. In this paper an extension to the DMP framework is introduced which facilitates the generation of trajectories for passive compliant link drives using reinforcement learning. The compliance of the joint actuators is preserved entirely during motion as well as at the goal position. The proposed approach is evaluated in simulation using a simplification of the common articulated configuration with 2 DOF and passive compliant link drives. Experiments are presented for point-to-point movements and different load case scenarios. The results demonstrate that the proposed approach is capable of generating trajectories for point-to-point movements to trained and untrained goal positions as well as for trained and untrained load cases.