Our recent paper, Learning from Trajectories via Subgoal Discovery, in NeurIPS 2019 proposes an approach to discover subgoals to solve complex tasks, from a limited number of expert demonstrations. Using these discovered subgoals leads to significantly faster robotic task learning with none to limited hand-designing of rewards. This work is done in collaboration between UCR and MERL (Mitsubishi Electric Research Laboratories). 

Title: Learning from Trajectories via Subgoal Discovery

           S. Paul, Jeroen van Baar, A. Roy-Chowdhury, Advances in Neural Information Processing Systems (NeurIPS), 2019