TechPrufer
Get started
NVIDIA
NVIDIA logoNVIDIA · Feature trail

Closed-Loop Autonomous Vehicle Model Training

1 post · Jun 2026 – Aug 2026

What this trail covers

This feature thread tracks the evolution of closed-loop training for autonomous vehicle (AV) models, enabling them to learn from the consequences of their own actions in simulation. Initial efforts focused on bridging the gap between open-loop training and real-world deployment by introducing frameworks like NVIDIA Alpamayo, AlpaSim, and AlpaGym. This post details the process of using AlpaGym to post-train AV models in a closed-loop reinforcement learning setup, connecting simulator feedback directly to the policy training loop. It covers installation, reward definition, and launching training, extending the capability to refine policies based on simulated environmental interactions.

All NVIDIA trails
  1. Generate Trajectories, Reasoning Traces, and Auto-Labels with NVIDIA Alpamayo 2 Super | NVIDIA Technical BlogIntroduces NVIDIA Alpamayo 2 Super, a 34B parameter open reasoning vision-language-action (VLA) model that jointly generates trajectories, Chain-of-Causation (CoC) reasoning traces, meta-actions, and grounded VQA outputs for autonomous vehicle (AV) development. It extends previous work by integrating a 32B Cosmos 3 Super Reasoner with a 2B Action Expert, post-trained with reinforcement learning. The model supports 360-degree perception from up to seven cameras and achieves state-of-the-art results in trajectory prediction (minADE_6 of 0.911 m), AV reasoning (0.433), and LingoQA (79.2). The post details workflows for generating these outputs and highlights their utility across the AV development lifecycle, including offline policy teaching, evaluation, data curation, and task customization, with integration into AlpaSim for closed-loop simulation.
  2. How to Post-Train Autonomous Vehicle Models in Closed-Loop with NVIDIA Alpamayo | NVIDIA Technical BlogThis post details the practical steps for using NVIDIA AlpaGym to perform closed-loop post-training of autonomous vehicle models. It outlines the installation process for AlpaGym, including host dependencies like CUDA, cuDNN, NCCL, and Redis, as well as Hugging Face authentication for dataset access. The post provides guidance on defining closed-loop rewards by combining AlpaSim metrics such as progress, collision avoidance, and off-road penalties. It also demonstrates how to launch training using AlpaGym's Hydra configuration, starting with a pre-trained Alpamayo model and converting it to an AlpaGym-compatible checkpoint. The workflow emphasizes connecting simulator feedback directly to the policy training loop for reinforcement learning, enabling models to learn from the consequences of their actions in simulation.