NVlabs/alpamayo

NVIDIA Alpamayo 1 Nano is an open 10B reasoning VLA model for autonomous vehicles that pairs driving trajectories with Chain-of-Causation reasoning.

What it solves

Alpamayo 1 is designed to improve the generalizability of autonomous driving, specifically addressing "long tail" scenarios where standard models often struggle. It bridges the gap between high-level reasoning and concrete action prediction (trajectory planning) to help vehicles navigate complex environments more reliably.

How it works

The project uses a Vision-Language-Action (VLA) architecture combining a Cosmos-Reason backbone with an action expert. It takes multi-camera video and egomotion history as inputs to produce two outputs: a reasoning trace (using a method called Chain-of-Causation reasoning) and a predicted trajectory consisting of 64 waypoints over a 6.4-second horizon at 10 Hz.

Who it’s for

This tool is intended for researchers and developers working on autonomous vehicle (AV) stacks, specifically those looking for a foundation model to build customized AV applications or create reasoning-based auto-labeling tools.

Highlights

  • Chain-of-Causation (CoC) reasoning: Uses hybrid auto-labeling with human-in-the-loop to generate reasoning traces.
  • VLA Architecture: Integrates vision, language, and action prediction into a single model.
  • Trajectory Prediction: Predicts precise waypoints for a 6.4s horizon.
  • SFT and RL Pipelines: Includes supervised fine-tuning (SFT) and reinforcement learning (RL) post-training pipelines via Alpamayo Recipes.