Core Concepts
Introducing FUSION, a safety-aware structured scenario representation method in offline RL for enhancing the safety and generalizability of autonomous driving agents.
Abstract
In the domain of autonomous driving, offline Reinforcement Learning (RL) approaches are effective but face challenges in maintaining safety. FUSION leverages causal relationships to enhance safety and generalizability. Extensive evaluations show improvements over current safe RL and IL baselines. The method significantly enhances safety and generalizability, even in challenging environments. Ablation studies confirm the benefits of integrating causal representation into the offline safe RL algorithm.
Stats
FUSION significantly enhances safety and generalizability compared to current state-of-the-art safe RL and IL baselines.
Empirical evidence attests to noticeable improvements with causal representation integration into offline safe RL algorithm.