How Do You Scale a Robotaxi Fleet to Six Million Vehicles?
NVIDIA has laid out the computing stack that every major robotaxi program now uses to move from prototype to commercial service. The platform spans three separate computers — one for training, one for simulation, and one inside each vehicle — and it is already running in fleets across four continents.
The Three Computers That Run a Fleet
A robotaxi stack is not a single box. It is three computers wired together by software that NVIDIA calls an open platform.
The first computer sits in a data center. NVIDIA DGX systems train the driving models that will eventually run in every car.
That second computer also lives in the cloud. It runs NVIDIA Omniverse and Cosmos on RTX PRO servers to simulate driving scenarios that real cars rarely encounter.
Inside the vehicle, the third computer rides along. DRIVE Hyperion 10 pairs two DRIVE AGX Thor chips built on the Blackwell architecture with a sensor suite that includes fourteen cameras, nine radars, three lidars, and twelve ultrasonic sensors.
Redundancy is the point. If one sensor or one chip fails, the car keeps driving.
Training Models That Reason Through Corner Cases
Training data grows as fleets drive more miles. The Alpamayo portfolio gives developers open reasoning models, simulation frameworks, and physical AI datasets they can adapt to their own data.
Reasoning models break complex situations into smaller steps. They select the safest trajectory after working through each step.
On a challenging autonomous driving evaluation, the addition of meta-action and chain-of-thought reasoning data cut the minimum average displacement error by forty-three percent. The error dropped from two point zero eight to one point one eight.
NVIDIA's physical AI datasets and reinforcement learning blueprints help developers optimize models for their target vehicles. Post-training recipes and distillation recipes are also provided.
Simulation That Turns Thousands of Miles Into Millions
Physical miles cannot capture every rare scenario. Omniverse NuRec models reconstruct real driving scenes from sensor data.
Cosmos world foundation models then generate physically based variations. They turn thousands of corner cases into millions of combinations covering behavior, traffic, weather, lighting, and sensor conditions.
Closed-loop simulation runs on RTX PRO servers. The AlpaSim framework extends the workflow for training and evaluating reasoning-based models.
De
The sources this article was written from
- NVIDIA Blog
https://blogs.nvidia.com/blog/robotaxi-leaders-full-stack-open-platform/