Learning an Explainable Trajectory Generator Using the Automaton Generative Network (AGN)

Symbolic reasoning is a key component for enabling practical use of data-driven planners in autonomous driving. In that context, deterministic finite state automata (DFA) are often used to formalize the underlying high-level decision-making process. Manual design an effective DFA can be tedious. combination with deep learning pipelines, serve as representation learn and process complex behavioral patterns. The goal this work leverage potential. We propose automaton generative network (AGN), differentiable DFAs. resulting neural network module standalone or embedded within larger architecture. evaluations on based vehicle planning tasks, we demonstrate incorporating AGN improves explainability, sample efficiency, generalizability model.