NVIDIA’s New AI Just Cracked The Hardest Part Of Self Driving

NVIDIA has released what appears to be the first open reasoning system for self-driving cars, including model weights, inference code, and a subset of training data. Unlike black-box systems, this AI explains its decisions in natural language before acting, reducing close encounter rates by 25%. Key techniques include reinforcement learning with a consistency reward (penalizing the model when actions don't match stated intentions), conditional flow matching for smoother motion, and training on 700,000 annotated video clips. A photorealistic simulator called Alpa Sim using 3D Gaussian splatting enables safe training on rare edge-case scenarios. The open release means researchers and students can now download and evaluate a state-of-the-art self-driving model independently of proprietary systems.