查看文章

The Driving SMARTS (Scalable Multi-Agent Reinforcement Learning Training School) competition was designed to address one of the major challenges for autonomous driving (AD), namely adaptation to distribution shift between data used for training and inference and the problems caused by this shift in real-world conditions. The two key features of the competition are 1) a two-track structure to encourage and support a variety of approaches to solving the problem, such as reinforcement learning, offline learning, and other machine learning methods; and 2) curated data for driving scenarios of varying difficulty levels, from cruising to unprotected turns at unsignalized intersections. The competition attracted 87 participants in 53 teams. Top-ranking teams contributed a diverse set of solutions highlighting the effectiveness of different methodologies on safe motion planning for AD. This paper provides an overview of the Driving SMARTS competition, discusses its organisational and design aspects, and presents the results, insights, and promising directions for future research.