Progressive stochastic motion planning for human-robot interaction

This paper introduces a new approach to optimal online motion planning for human-robot interaction scenarios. For a safe, comfortable, and efficient interaction between human and robot working in close proximity, robot motion has to be agile and perceived as natural by the human partner. The robot has to be aware of its environment, including human motions, in order to proactively take actions while ensuring safety, and task fulfillment. Human motion prediction constitutes the fundamental perception input for the motion planner. The prediction system, which is based on probabilistic movement primitives, generates a prediction of human motion as a trajectory distribution learned in an offline phase. The proposed stochastic optimization-based planning algorithm then progressively finds feasible optimization parameters to replan the motion online that ensures collision avoidance while minimizing the task-related trajectory cost. Our simulation results show that the proposed approach produces collision-free trajectories while still reaching the goal successfully. We also highlight the performance of our planner in comparison to previous methods in stochastic motion planning.