Integrating simulation with robotic learning from demonstration

Robots that co-habitat an environment with humans, e.g., in a domestic or an agricultural environment, must be capable of learning task related information from people who are not skilled in robotics. Learning from demonstration (LfD) offers a natural way for such communication. Learning motion primitives based on the demonstrated trajectories facilitate robustness to dynamic changes in the environment and task. Yet since the robot and human operator typically differ, a phase of autonomous learning is needed for optimizing the robotic motion. Autonomous learning using the physical hardware is costly and time consuming. Thus finding ways to minimize this learning time is of importance. In the current paper we investigate the contribution of integrating an intermediate stage of learning using simulation, after LfD and before learning using robotic hardware. We use dynamic motion primitives for motion planning, and optimize their learned parameters using the PI2 algorithm which is based on reinforcement learning. We implemented the method for reach-tograsp motion for harvesting an artificial apple. Our results show learning using simulation drastically improves the robotic paths and that for reach-to-grasp motion such a stage may eliminate the need for learning using physical hardware. Future research will test the method for motion that requires interaction with the environment. Proceedings 28th European Conference on Modelling and Simulation