NVIDIA Launches Alpamayo 2 Super Reasoning Model For Robotaxis

NVIDIA today introduced NVIDIA Alpamayo 2 Super, a 32-billion-parameter reasoning-based vision language action (VLA) model that extends the NVIDIA Alpamayo family of open AI models, simulation frameworks and physical AI datasets for safe, level 4 robotaxi development.

Alongside the model, the company announced new tools, models and agent skills that complete the pipeline from real-world data capture to closed-loop training and in-vehicle deployment, including NVIDIA AlpaGym, NVIDIA OmniDreams and new NVIDIA Omniverse NuRec models.

Alpamayo 2 Super helps accelerate autonomous vehicle (AV) development by eliminating the need to build key autonomy infrastructure from scratch. It enables humanlike perception, reasoning and action, and provides the interpretability needed for safety validation and regulatory collaboration.

“Alpamayo is the moment cars begin to safely reason, not just drive. Only NVIDIA makes available open models, simulation, real-world data and agent skills so the entire global robotaxi ecosystem can develop level 4 capabilities that understand edge cases, explain decisions, earn trust and scale safely to millions of vehicles.”

Jensen Huang, founder and CEO of NVIDIA

Inside Alpamayo 2 Super, NVIDIA’s open reasoning model

The NVIDIA Alpamayo family now scales from 10 billion to 32 billion parameters with Alpamayo 2 Super — going beyond trajectory generation to reason, plan and act across the full driving stack. With multitask capabilities spanning reasoning, auto-labeling, scene understanding, model critiquing and distilling knowledge into smaller models, it provides the building blocks for scalable L4 AV development and deployment.

• 3x parameter scale: Built on NVIDIA Cosmos world foundation models, Alpamayo 2 Super scales to 32 billion parameters compared with previous 10-billion-parameter generations, improving reasoning, 3D spatial understanding and trajectory prediction in long-tail scenarios.
• Full-surround perception: Expands from front-focused cameras to 360-degree situational awareness across front, side and rear views, giving the model complete context for safer lane changes, merges and intersection crossing.
• Meta-Actions: Adds Meta-Action output — including macro actions such as yield, lane change and stop — so the model predicts high-level driving decisions for downstream planning in addition to trajectories and chain-of-causation (CoC) traces.
• Reasoning auto-labeling and 2D grounding: Introduces reasoning auto-labeling with 2D grounding so the 32-billion-parameter foundation model can provide high-quality reasoning labels, compressing annotation cycles from months to days and reshaping AV data pipeline economics.
• Improved CoC and trajectory quality: Improved CoC traces and trajectories, especially in rare, complex, long-tail scenarios where traditional imitation-learning AV stacks struggle.

These advancements make Alpamayo 2 Super NVIDIA’s most powerful open driving foundation model to date. Designed as a teacher model, Alpamayo 2 Super can be distilled into compact models that run on the accelerated compute of the NVIDIA DRIVE Hyperion platform — NVIDIA DRIVE AGX Thor, which runs inside the vehicle.

Alpamayo was recently recognized by the COMPUTEX Best Choice Awards, winning in the Vehicle Technology and Smart Cockpit category. Since launch, Alpamayo has been downloaded close to 400,000 times. The Alpamayo open platform also includes post-training scripts that allow researchers and developers to adapt the models to their own datasets, scenarios and driving policies. Alpamayo 2 Super is expected to be available this summer on GitHub for inference code and Hugging Face for model weights.

Closed-loop training: AlpaGym, OmniDreams and the reasoning model toolchain

NVIDIA also introduced NVIDIA AlpaGym, an open source, high-throughput, closed-loop RL framework. Where open-loop training evaluates models against recorded data and generates a single round of actions, AlpaGym runs models through continuous decision and observation cycles in NVIDIA AlpaSim, with every braking, steering and navigation choice affecting the environment. As a result, AlpaGym exposes the compounding errors and edge-case failures that static datasets miss, allowing models to learn from experience.

Built on the AlpaSim microservice simulation stack and NVIDIA Omniverse NuRec, AlpaGym enables efficient, scalable, closed-loop RL to push the frontier of driving performance. In combination with the Physical AI AV Dataset, Alpamayo provides a continuous path from open-loop pretraining to closed-loop refinement.

NVIDIA is also releasing the CoC Auto-Labeling Pipeline as open source on GitHub. The pipeline automatically generates decision-grounded and causally linked CoC labels from raw driving clips with no human annotation required, providing the causal training data foundation needed to train embodied reasoning models at scale.

To support reasoning-based AV development, NVIDIA is launching new physical AI agent skills, under NVIDIA Agent Toolkit, to guide developers and their coding agents through the simulation, data generation and closed-loop training workflows needed to build and validate autonomous driving systems at scale. This includes Neural Reconstruction skills powered by NVIDIA Omniverse NuRec libraries, NVIDIA OmniDreams skills for photorealistic scenario generation and AlpaGym skills for closed-loop RL.

For more autonomous driving news, click here.

Source