Do As You Teach: A Multi-Teacher Approach to Self-Play in Deep Reinforcement Learning

Chaitanya Kharyal, Tanmay Kumar Sinha, Sai Krishna Gottipati, Fatemeh Abdollahi, Srijita Das, Matthew E. Taylor

Abstract: The future of industrial automation is hinged on the ability of the industrial robots to precisely finish the tasks designated for them [5]. These tasks are usually specified in terms of a state the robot is required to reach (i.e., a goal state). Goal-conditioned reinforcement learning [7, 8] is an emerging sub-field that trains policies with goal inputs. This enables the agent to generalize to new unseen goals, learn multiple complex tasks and acquire new skills along the way.