Expanding the Alpamayo Open Platform for Developing Reasoning AVs Across Models, Data, and Simulation
NVIDIA today announced a significant expansion of the Alpamayo open platform for developing reasoning-based autonomous vehicles (AVs). First introduced at CES 2026, the Alpamayo open platform provides researchers and developers with a flexible, high-performance, and scalable suite of models, datasets, and simulation tools for evaluating and training modern reasoning-based AV stacks in realistic closed-loop settings.
Since its launch, the Alpamayo open platform has seen rapid adoption across both industry and academia, with more than 100,000 AV researchers and developers worldwide downloading the Alpamayo 1 AV reasoning model. Based in part on feedback from the community, we are introducing several enhancements to the Alpamayo platform across models, data, and simulation. Collectively, these additions enable developers to build AV stacks with more powerful reasoning models, evaluate them more flexibly, and expand their datasets with additional reasoning-focused data.
From NVIDIA Alpamayo 1 to NVIDIA Alpamayo 1.5: Making Reasoning VLAs More Powerful and Customizable
Alpamayo 1.5 is a significant update to NVIDIA’s open 10B-parameter chain-of-thought reasoning VLA model, designed to be an interactive and steerable reasoning engine for the AV community. Alpamayo 1.5 is built on the Cosmos-Reason2 VLM backbone, is RL post-trained, and introduces support for navigation guidance, flexible multi-camera counts, and user question answering. We are also releasing post-training scripts to enable model adaptation for researchers and developers.
✨ Key Highlights of this New Release
Text-guided trajectory planning: condition the model with natural-language driving commands (e.g., "turn left in 200m") to steer trajectory generation, enabling controllable and interpretable planning.
Flexible multi-camera support: work with a variable number of cameras without being locked to a fixed sensor rig, making it straightforward to apply the model to different vehicle platforms and camera layouts.
State-of-the-art performance in multiple aspects (reasoning quality, trajectory accuracy, alignment, and more), with particular improvements observed due to RL post-training.
Easy-to-use scripts and notebooks that enable application across a wide range of use cases, from fine-tuning with new data to general scene understanding.
Supervised fine-tuning (SFT) scripts: ready-to-run scripts for fine-tuning the Alpamayo 1 (and soon Alpamayo 1.5) model on your own driving data, enabling rapid domain adaptation.
Reinforcement learning (RL) post-training scripts: ready-to-run scripts for fine-tuning the Alpamayo 1 (and soon Alpamayo 1.5) model with customized rewards to improve reasoning, trajectory quality, and alignment between these two output modalities, as well as with desired driving behaviors.
🤖 Popular Use Cases of Alpamayo 1.5
AV model distillation – Leverage the pretrained weights as an offline teacher to develop onboard-ready models (e.g., via output or feature supervision during training).
Data labeling and curation - Identify interesting scenarios and label them with plausible future trajectories and reasoning traces.
Model customization - Post-train the model with your own data and labels, specialize Alpamayo to best suit your needs.
Visual question answering – Ask specific questions about scenes, leverage the outputs for data curation or autolabeling.
AV evaluation – Generate trajectories and reasoning traces to evaluate the outputs of smaller, edge-deployed models, or assess alternative outcomes by providing different navigation commands to the Alpamayo 1.5 model.
Reasoning Through the NVIDIA Physical AI AV Dataset
The PhysicalAI-Autonomous-Vehicles dataset provides one of the largest, most geographically diverse collections of multi-sensor data for AV researchers to build the next generation of reasoning-based end-to-end driving systems. Today, we’re announcing the release of reasoning labels for the dataset, enabling developers to finetune their own reasoning models. We are also releasing our chain-of-causation autolabeling pipeline. The associated physical_ai_av developer kit makes it easy to get started immediately with the dataset, containing a data interface and a detailed wiki about the format and structure of the dataset.
✨ Key Highlights of this New Release
High-quality Reasoning labels: We’re releasing manually verified reasoning labels for a subset of the Physical AI AV dataset, enabling researchers to develop and evaluate their own reasoning models with high-quality labels.
(Coming Soon) CoC autolabeling pipeline: To bring the benefits of chain-of-causation (CoC; a more powerful way of representing reasoning traces) to all, we will release an automated workflow that generates reasoning labels for driving data.
(Coming Soon) Out-of-Distribution Benchmark: We will release a PhysicalAI-AV Out-of-Distribution (PAI-OOD) Benchmark to evaluate how accurately models reason through difficult scenarios, ranging from semantic anomalies (e.g., rare agent types, behaviors), conditions (e.g., weather, sensor noise), rare behaviors (e.g., erratic motion, illegal driving), and other uncommon events.
🤖 Popular Use Cases of the PhysicalAI-AV Data and Labeling Pipeline
AV model training and evaluation – Train and evaluate your end-to-end AV model on data from a variety of geographies and driving conditions, with a focus on rare events.
Reasoning autolabeling – Add reasoning labels to your data, train with them to elicit reasoning capabilities from your desired model(s).
Neural reconstruction and rendering – Leverage multi-sensor data to create and evaluate new neural reconstruction and rendering techniques.
Anomaly detection and safety evaluation – Explore how your model handles unusual events by evaluating it on unseen geographies, agent types, and behaviors.
NVIDIA AlpaSim: Even More Flexible Closed-Loop AV Simulation
AlpaSim is an open-source end-to-end AV simulation platform designed specifically for research and development. It enables users to test end-to-end AV policies in a closed-loop setting by simulating realistic sensor data, vehicle dynamics, and traffic scenarios within a modular and extensible testbed. Today, AlpaSim is even more flexible and extensible with a new microservice plugin system, and the concurrent GA release of NuRec enables evaluation on a much wider variety of driving data (in addition to the 900+ reconstructed scenarios already available on Hugging Face).
✨ Key Highlights of this New Release
The amount of scenes that can be used with AlpaSim is greatly expanding with the GA release of NuRec.
Higher quality rendering and support for more camera types via NuRec 26.02.
A new extensible plugin system allows AlpaSim to more easily interact with additional microservices (new drivers, renderers, data sources, etc).
We will be launching an AlpaSim benchmark, leaderboard, and challenge at academic conferences to advance the development of state-of-the-art closed-loop driving models. Stay tuned for more information and announcements!
🤖 Popular Use Cases of AlpaSim
Algorithm testing – Test new autonomous driving algorithms with our dataset, your own desired environments, and/or your own simulation components.
Safety analysis – Evaluate vehicle behavior in edge cases and challenging scenarios.
Performance benchmarking and regression testing – Compare different models and configurations, e.g., accuracy under different camera setups.
Debugging – Understand and debug complex autonomous driving behavior.
Conclusion
Reasoning models in AVs will unlock new capabilities and levels of safety for the next generation of autonomous systems. We hope that this update, and future Alpamayo releases, will accelerate the whole industry with new resources and tools.
Resources
For more information about this release, please check the below resources (please note that most of the links will become active in the next few days).
Alpamayo Models:
- Alpamayo 1.5 Model Weights → https://huggingface.co/nvidia/Alpamayo-1.5-10B
- Alpamayo 1.5 Inference Code → https://github.com/NVlabs/alpamayo1.5 (with post-training scripts to follow shortly)
- Alpamayo 1 Post-Training Scripts → https://github.com/NVlabs/alpamayo
Physical AI AV Dataset:
- Dataset (with new reasoning labels) → https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles
- Developer Kit → https://github.com/NVlabs/physical_ai_av
- CoC Autolabeling Pipeline → coming soon
AlpaSim:
- Code → https://github.com/NVlabs/alpasim
- Reconstructed Scenarios → https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles-NuRec
- NuRec GA Release → https://docs.nvidia.com/nurec/
Research on Reasoning VLA Models, End-To-End AV Simulation and Training, and Physical AI Safety:
- Website → https://research.nvidia.com/labs/avg/
