AlpamayoR1: Large Causal Reasoning Models for Autonomous Driving

Nvidia's AlpamayoR1 is a new architecture for autonomous driving that uses a large Vision-Language Model (VLM) for causally-grounded reasoning. A core innovation is the "Chain of Causation" concept, which leverages a specially annotated dataset to train the model on causally consistent driving scenarios. The system processes visual inputs via a Vision Transformer and generates reasoning traces that inform a trajectory decoder, which uses flow-matching for efficient inference. The model's performance is refined through a multi-stage process involving supervised fine-tuning and reinforcement learning post-training to improve safety and consistency.