Learning to Plan with Logical Automata

We present a framework for combining imitation learning and logical automata and introduce the Logic-based Value Iteration Network (LVIN) model. By appending a ‘logical’ dimension to the state space of the environment, LVIN can recover and incorporate the transition matrix of a finite state automaton derived from a Linear Temporal Logic formula into the policy learned through imitation. This approach improves upon the original VIN in several ways: LVIN (1) is capable of learning logic corresponding to long sequences of steps, (2) adapts easily to new specifications, and (3) is amenable to correction after faulty expert demonstrations (e.g., in a driving domain).