Nvidia says major robotaxi programmes use its stack

Nvidia says every major robotaxi programme operating at commercial scale uses parts of its autonomous vehicle technology stack, which spans AI training, simulation, validation and in-vehicle computing.

The claim places Nvidia at the centre of a fast-growing robotaxi market that could reach USD $400 billion by 2035, with more than 6 million commercial vehicles in operation.

It described robotaxi development as a three-computer system covering model training in the data centre, simulation and validation in virtual environments, and real-time computing inside vehicles.

For training, Nvidia highlighted its DGX systems and Alpamayo portfolio of vision, language and action models, simulation frameworks and physical AI datasets. Developers can adapt those tools to their own data and software. Nvidia added that reasoning models are meant to handle difficult edge cases in autonomous driving by breaking decisions into smaller steps.

To support that pitch, Nvidia cited internal testing. In one autonomous driving evaluation, it said adding meta-action and chain-of-thought reasoning data improved trajectory prediction accuracy and reduced minimum average displacement error by 43%, from 2.08 to 1.18.

Simulation stack

Simulation remains central for robotaxi operators because real-world driving alone does not produce enough examples of rare or dangerous scenarios. Nvidia said its Omniverse NuRec models can reconstruct real-world driving scenes from sensor data, while its Cosmos world foundation models can generate variations in traffic, weather, lighting and sensor conditions.

The tools are intended to help developers turn thousands of corner cases into much larger sets of test conditions. Nvidia said the software runs on RTX PRO Servers and supports closed-loop simulation and validation, while its AlpaSim framework is used to train and evaluate reasoning-based driving models before deployment on roads.

Inside the vehicle, Nvidia highlighted DRIVE Hyperion 10, its reference architecture for level 4-ready robotaxis. The system combines dual DRIVE AGX Thor chips with a sensor suite that includes 14 cameras, nine radars, three lidars and 12 ultrasonic sensors.

According to Nvidia, the design includes redundancy in both computing and sensing so the vehicle can continue operating if one element fails. It also said its Halos safety system provides an operating system and a broader validation framework covering inspection, validation, simulation and continuous testing.

Industry uptake

Nvidia used the announcement to list a broad range of partners and customers across North America, Europe, Asia and the Middle East, spanning ride-hailing groups, autonomous driving developers, vehicle manufacturers and component suppliers.

Among the most prominent names, Uber is scaling a fleet based on DRIVE Hyperion and working with Nvidia on a robotaxi AI data factory that uses Cosmos to curate driving data for rare scenarios. Uber is also collaborating with Autobrains, Avride, Lucid, May Mobility, Mercedes-Benz, Momenta, Nissan, Nuro, Pony.ai, Stellantis, Waabi, Wayve, WeRide and Zoox.

Other transport platforms are part of the push as well. Bolt is using Nvidia technology in its autonomous vehicle efforts in Europe, while Lyft plans to use DRIVE Hyperion as a reference architecture for future autonomous fleets. May Mobility vehicles already operate on Lyft's network in Atlanta using Nvidia DRIVE.

Several autonomous vehicle developers are building their software and vehicle systems around Nvidia hardware and software. Wayve, Nissan and Uber are working on a global robotaxi programme using a prototype vehicle that combines Nissan engineering, Wayve's embodied AI and DRIVE Hyperion. Zoox uses Nvidia DRIVE for in-vehicle computing as well as cloud-based training and simulation, while Pony.ai developed its new autonomous driving domain controller with DRIVE Hyperion and DRIVE AGX Thor.

Momenta is developing its software stack on DRIVE AGX running on DriveOS. Tensor is building its level 4 Robocar with eight DRIVE AGX Thor systems-on-a-chip in its in-vehicle supercomputer, and Waabi's Waabi Driver platform is built on DRIVE AGX Thor as it moves into robotaxi deployment with Uber.

Production phase

Automakers are also moving beyond development towards production programmes that include Nvidia technology. Tesla uses Nvidia supercomputers to train its autonomous-driving neural networks, while Mercedes-Benz is working with Nvidia and Uber on a robotaxi ecosystem based on the S-Class and Nvidia's level 4 software and AI models.

Stellantis, Wayve and Uber are also developing driverless mobility services using DRIVE Hyperion and Nvidia AI computing products. Lucid, Nuro and Uber are working on a global robotaxi service using DRIVE AGX Thor, and Hyundai Motor and Kia are expanding their collaboration with Nvidia on data-driven autonomous-driving systems built on DRIVE Hyperion.

Elsewhere, Geely and its partners plan to develop and commercialise robotaxis using DRIVE Hyperion, while Zeekr has adopted DRIVE AGX Thor for a centralised domain controller. Lenovo is supplying an AD1 level 4 domain controller based on DRIVE AGX Thor for a next-generation robotaxi programme with SWM.

The breadth of those relationships shows how robotaxi economics increasingly depend on a common set of computing tools, from training data centres to simulation software and vehicle hardware. Rather than backing a single operator, Nvidia is positioning itself as a supplier across nearly every layer of the robotaxi market, from cloud systems used to train models to the processors and sensor architectures installed in the vehicles.

From cloud to car, nearly every layer of the robotaxi platform is being developed on Nvidia accelerated computing.