Learning Human-like Trajectories for Whole-Body Motion with Artificial Force Fields
Presumably in hybrid societies, robots will only be accepted if their motions are predictable and human-like. Therefore, the project aims to optimize robot trajectories in terms of human-likeness. New methods will be developed for generating collision-free and functional trajectories by applying an end-to-end learned deep neural network, which uses depth maps. Deep neuronal networks are trained autonomously by traditional sampling-based motion planning methods. Moreover, a deep reinforcement learning approach adapts robot trajectories with artificial force fields such that they increase human-likeness. Several studies will be conducted to obtain a useful scale for human acceptance.