GPUDrive: Data-driven, multi-agent driving simulation at 1M FPS

The paper titled "GPUDrive: Data-driven, multi-agent driving simulation at 1 million FPS" presents a new GPU-accelerated simulator designed to enhance multi-agent planning through rapid experience generation. The authors, Saman Kazemkhani and colleagues, highlight that traditional multi-agent learning algorithms require extensive experience, often in the billions of steps, which has limited their practical application. GPUDrive, built on the Madrona Game Engine, addresses this bottleneck by generating over a million steps of experience per second. The simulator allows for the definition of complex agent behaviors using C++, which are then optimized for performance with CUDA. The authors demonstrate the effectiveness of GPUDrive by training reinforcement learning agents on the Waymo Motion dataset, achieving rapid training times for both individual scenes and general capabilities. The trained agents are made available as part of the code base, promoting further research and development in multi-agent systems.

- GPUDrive can generate over a million steps of experience per second for multi-agent simulations.

- The simulator is built on the Madrona Game Engine and utilizes C++ and CUDA for performance optimization.

- It addresses the challenge of extensive experience requirements in traditional multi-agent learning algorithms.

- The authors successfully trained reinforcement learning agents using the Waymo Motion dataset.

- The trained agents are included in the code base for public access and further research.