NVIDIA officially launched the Alpamayo 2 Edge in-vehicle AI model on November 11, emphasizing its real-time adaptive reasoning capabilities. The model can plan driving routes based on the surrounding environment and explain its driving decisions when necessary. For example, when encountering pedestrians, the system will proactively yield, or it will make space for oncoming buses on narrow roads. NVIDIA stated that developers can build next-generation autonomous vehicles based on this open model and customize it according to their fleet needs.
Alpamayo 2 Edge employs approximately 2 billion parameters in its Nemotron reasoning backbone network, combined with a distributed action expert, and belongs to the visual-language-action (VLA) model category. It can simultaneously process multi-camera images around the vehicle, historical motion information, and navigation instructions to generate driving trajectories, while using causal reasoning to explain why a particular driving action was taken. Unlike traditional autonomous driving systems that primarily rely on preset rules and perception modules, this approach aims to give vehicles a more human-like understanding of scenes and decision-making capabilities.
Significant Improvement in Adaptive Reasoning Capabilities of Alpamayo 2 Edge
The focus of this upgrade is adaptive reasoning. When driving on normal roads, the vehicle does not need to perform complex reasoning at every moment; the model can reduce unnecessary computations. However, when faced with complex road conditions, it can generate a more detailed reasoning process. This design reduces the consumption of in-vehicle computing resources while retaining decision-making capabilities in complex scenarios.
In terms of performance, NVIDIA trained the model using the AlpaGym closed-loop reinforcement learning framework, resulting in a 48.9% improvement in closed-loop driving performance compared to the best model that did not undergo this training phase. After optimization with tools such as TensorRT, the inference latency for image-to-driving trajectory on the DRIVE AGX Thor platform is approximately 160 to 190 milliseconds. Developers can also download the model weights to fine-tune according to their vehicle and sensor configurations, utilizing NVIDIA's simulation, training, and evaluation tools to optimize driving strategies.
Alpamayo 2 Edge is currently primarily aimed at autonomous driving developers and is not a complete autonomous driving system that can be directly installed in standard production vehicles.

