Alpamayo 2 Super from NVIDIA is a model that develops solutions for robotaxis.

The new development is aimed at robotaxis and is designed to function as a unified intelligent layer that integrates environmental understanding, motion planning, and real-time action selection.

What is Alpamayo 2 Super

Alpamayo 2 Super is a vision-language-action model with 32 billion parameters. It is designed to enable autonomous vehicles to not just "see" the road, but to interpret events and make decisions based on context.

The model processes visual data, understands the traffic situation, forms a behavior plan, and outputs actions that can be used by the vehicle's control system. This makes it an integral part of the full autonomous driving cycle, where all stages are integrated into a single architecture.

Scaling and Enhanced Understanding

Compared to previous versions of the Alpamayo family, the new model has been significantly scaled up. Previously, models with around 10 billion parameters were used; now, the size has increased to 32 billion.

This scaling aims to improve the model's ability to handle complex traffic scenarios, better understand spatial relationships between objects, and more accurately predict the evolution of situations on the road. This is particularly important in dense urban traffic and rare, non-standard situations.

Full 360-Degree Perception

A key update is the expansion of environmental perception to a full circular view. The model now uses data not only from front-facing cameras but also from side and rear cameras.

This allows the system to form a more holistic representation of the traffic situation. This helps during lane changes, intersections, and any scenarios where it is important to account for objects outside the direct forward field of view.

Meta-Actions and Transition to High-Level Decision Making

Alpamayo 2 Super introduces the concept of Meta-Actions. This marks a shift from simple trajectory prediction to higher levels of decision-making.

The model can generate actions such as stopping the vehicle, changing lanes, or yielding to other traffic. This makes the system's behavior more structured and closer to human driving logic, where semantic understanding of the situation is as important as movement coordinates.

Accelerated Learning Through Automated Labeling

NVIDIA has added new learning approaches to the system, significantly speeding up data preparation.

One such mechanism is reasoning auto-labeling, where the model helps automatically generate labels based on the logic of events in the scene. Additionally, 2D grounding is used to link the model's reasoning to specific objects in the image.

Thanks to these approaches, the data preparation and labeling process is significantly accelerated.

AlpaGym — Learning Through Simulation

Alongside the model, NVIDIA introduced AlpaGym, a system for training in a closed-loop simulation environment.

In this environment, the model can test its actions, observe consequences, and adjust its behavior based on errors. This approach allows the system to be trained not only on static data but also on dynamic interactions with a virtual road environment.

OmniDreams and Generation of Complex Scenarios

To expand training data, the OmniDreams system is used. This is a generative model that creates realistic traffic scenarios, including rare and complex situations.

It allows for the modeling of the so-called "long tail" of scenarios that rarely occur in real life but are critically important for the safety of autonomous systems. This makes training more comprehensive and robust against non-standard situations.

Physical AI Agent Skills and Development Tools

NVIDIA has also expanded the set of tools for developers of autonomous systems. These include technologies for reconstructing real-world data, generating synthetic scenes, and closed-loop training.

All these tools are integrated into a single ecosystem that supports the development of physical artificial intelligence and enables the creation of more realistic behavior models for autonomous vehicles.

The Model's Role in the NVIDIA Ecosystem

Alpamayo 2 Super is positioned as a teacher model that can be used to transfer knowledge to compact systems. It serves as a teacher model and can be applied for distillation into automotive computing platforms.

The model will be available in the summer as open-source inference code on GitHub and weights on Hugging Face, making it part of the open ecosystem for autonomous vehicle development.