Perspectives on Robot Learning: Imitation and Causality


Organizers: Animesh Garg, Michael Laskey, Yuke Zhu, Jiajun Wu and Stefano Ermon

Website: https://sites.google.com/stanford.edu/rss18-causal-imitation

Sequential Decision Making and Reinforcement Learning in complex environments with sparse rewards and stochastic dynamics is a long standing challenge. Despite the recent success of RL in games, there are challenges in applying these methods to robotics in the face of safety concerns and cost of environmental interaction. At the same time absence of an informative reward function can render this family of iterative learning methods impractical. In contrast, Imitation learning algorithms guide an agent towards the correct behavior via leveraging a supervisor. Imitation may imply "do the same thing"; but ideally we seek to emphasize semantic similarity rather than literal behavior cloning. Often such generalization may require exploration that goes beyond trajectory replay. For instance, a robot may replicate a human trajectory to open the door but might fail to opening a window or a fridge. Such generalization needs a representation of the task that facilitates causal exploration. We need to build joint action-perception representations that encode perceivable effects, and select actions in terms of operations that determine intended future percept from the given current percept.Recent research has reiterated the efficiency of imitation learning based methods over RL for learning in physical domains as well as addressing problems of limited non-i.i.d. data in Imitation. At the same time research in causality has resulted in promising abstractions for robotics. There is an exciting opportunity in combining these ideas to achieve generalization -- whereby imitation guides task representations, and causality enables exploration for generalization. This workshop will serve as a platform to discuss the impact and merit of algorithmic techniques in Imitation Learning and Causal Inference, and their applications in robotics. We invite submissions advancing the theory, abstractions and systems in both imitation and causality for robotics.