From Sampling to Learning: Accelerating Human-Centered Robot Motion Generation

The increasing deployment of robots in shared spaces poses challenges for motion planning and demands reactive, human-aware motions rather than purely functional trajectories. In such scenarios, motion appearance and planning delay significantly impact collective task performance and the overall acceptance of a robotic system. This thesis examines how collision-free robot motions can be generated rapidly to facilitate human-robot collaborations. The first part examines how characteristics of human arm motions can be integrated into robot motion planning using a human-likeness cost function. Next, an extensive user study examines the effect of such motions on their interpretability and the perceived naturalness of the robotic system. The second part addresses long planning delays that often occur in complex environments and can hamper human-robot interaction. To this end, three approaches are proposed. First, a generative neural network is integrated into an optimizer to solve the dual-arm inverse kinematic problem. Second, hyperparameter optimization is used to correctly parameterize motion planners for the problem at hand. Finally, a generative neural sampler is proposed that integrates seamlessly into state-of-the-art sampling-based algorithms. The resulting neural motion planner reduces planning delay drastically and enables rapid global motion planning. In the third part, dual-arm robot motion planning is addressed as an imitation learning problem. For this, two motion-planning policies are trained on an extensive dataset of expert motions to generate joint-space trajectories. This work thus provides insights into how recent advancements in imitation learning transfer into the domain of dual-arm robot motion planning.