Maximising Coefficiency of Human-Robot Handovers through Reinforcement Learning

Handing objects to humans is an essential capability for collaborative robots. Previous research works on human-robot handovers focus facilitating the performance of human partner and possibly minimising physical effort needed grasp object. However, altruistic robot behaviours may result in protracted awkward motions, contributing unpleasant sensations by affecting perceived safety social acceptance. This paper investigates whether transferring cognitive science principle that "humans act coefficiently as a group" (i.e. simultaneously maximising benefits all agents involved) cooperative tasks promotes more seamless natural interaction. Human-robot coefficiency first modelled identifying implicit indicators comfort discomfort well calculating energy consumption performing desired trajectory. We then present reinforcement learning approach uses score reward adapt learn online combination interaction parameters maximises such coefficiency. Results proved acting could meet individual preferences most subjects involved experiments, improve comfort, foster trust robotic partner.