Automata Embeddings for Goal-Conditioned Reinforcement Learning

This talk by Beyazit Yalcinkaya from UC Berkeley explores a novel approach to goal-conditioned reinforcement learning (GCRL) using automata as goal representations. Learn how automata can provide formal temporal semantics while remaining interpretable, similar to flow charts. The presentation addresses the challenges of conditioning agent behavior on automata, which form an infinite concept class with Boolean semantics where small changes can significantly alter tasks. Discover a technique for learning provably correct embeddings of "reach-avoid derived" automata that guarantees optimal multi-task policy learning. The speaker demonstrates through empirical evaluation how the proposed pretraining method enables zero-shot generalization across various task classes and accelerates policy specialization without the myopic suboptimality typically found in hierarchical methods.