Inside NVIDIA Alpamayo 1: Making Autonomous Vehicles Reason
In this livestream, we’ll take a deep dive into Alpamayo 1 — a 10B-parameter, chain-of-thought reasoning VLA model that processes multi-camera video and outputs both driving trajectories and reasoning traces, revealing the logic behind each decision. We’ll cover key design choices, highlight main results, and discuss future directions, as well as answer questions developers may have about applying it to their own use cases. Key takeaways: - Understand how chain-of-thought reasoning VLA models like Alpamayo 1 are designed and built. - Explore how Alpamayo 1 can be applied across use cases, from powering end-to-end autonomous driving backbones to enabling reasoning-driven auto-labeling tools. - Learn about opportunities and challenges in the development of reasoning models for Physical AI. Get started with the broader Alpamayo ecosystem: https://huggingface.co/blog/drmapavone/nvidia-alpamayo ---------------------------------------------------------------------------- Key Moments: 0:02 Welcome and introduction to Alpamayo 1 1:31 Alpamayo ecosystem overview 2:53 Video overview: reasoning-driven autonomy 4:48 Alpamayo pillars: model, simulator, dataset 5:44 Foundation model principles for end-to-end AV 7:15 Language pre-training improves trajectory quality 9:01 From end-to-end driving to reasoning-based models 10:48 What is a chain of causation reasoning trace 11:21 Multi-stage training pipeline overview 13:31 Action representation design for driving 15:07 Discrete tokens vs diffusion for trajectory generation 17:44 Structured chain of causation auto-labeling 20:34 Supervised fine-tuning with reasoning traces 22:30 Embodied misalignment and post-training alignment 25:03 Reinforcement learning for reasoning-action consistency 27:44 Closed-loop simulation results in AlpaSim 28:37 Real-world deployment and distilled models 29:51 Scaling laws: model size and data 30:55 Alpamayo roadmap: Nano to Super to Ultra 33:41 Live demo: downloading and running Alpamayo 1 37:56 Prompt structure and input formatting 39:48 Generating reasoning traces and trajectories 41:19 Second scenario inference example 42:39 Community Q&A 1:00:25 Human driving logic and alignment discussion 1:01:00 Closing remarks and GTC preview ---------------------------------------------------------------------------- 👥Join the Community → https://nvda.ws/3ZMfc6e ---------------------------------------------------------------------------- #AutonomousVehicles #PhysicalAI #VisionLanguageModel #ReasoningModels #NVIDIA #Alpamayo #AlpaSim #SelfDriving #ReinforcementLearning #FoundationModels #EdgeAI #ClosedLoopSimulation #AIResearch