Predicting humans for robust handovers: design and robotic implementation of a learning model for online trajectory generation

Recent years have seen an increasing interest in collaborative manipulation tasks. In this context, robots should be capable of handing over objects for a successful cooperation with humans. The execution of a fluent pass, although representing an undemanding action that can be executed effortlessly among humans, is still an open challenge for robots, due to its underneath complexity. This work aims to develop a learning model that can allow a robotic system to execute a fluent handover without any a priori insight about the final point of exchange. To this end, an ongoing prediction of the human hand trajectory is provided to a minimum jerk tracking system, which determines the path the robot should follow to execute the handover. Moreover, an error metrics is computed to understand the human intent to perform the task, modulating the approaching speed of the robotic system. The system was tested in a real scenario. It was proven to be robust to perturbations, managing unexpected human movements and reaching the predicted target point coordinating with the partner