FAST: A Framework for Aligned Sampling and Training in Parallel Reinforcement Learning for Autonomous Driving

The article presents FAST, a framework designed to enhance sampling efficiency in parallel reinforcement learning for autonomous driving. FAST introduces Dynamic Parallel Sampling Alignment (DPSA) to address the straggler effect by extending terminated episodes and implementing global truncation based on termination rates, which allows for improved sample utilization without re-initialization delays. Empirical results show that FAST achieves a minimum of 1.78 times wall-clock speedup compared to single-clip baselines while maintaining statistical unbiasedness, making it a significant advancement for practitioners focused on optimizing reinforcement learning processes in autonomous systems.