“The goal-solving policy (Bob) is trained with a combination of a goal-conditioned reinforcement learning reward and behavior cloning using the goal-generating policy's (Alice's) trajectory as a demonstration.”