Lifelike Simulation Accelerates Middle-Mile Autonomous Trucking

The Gatik Arena platform integrates NVIDIA Cosmos models to create closed-loop, ultra-realistic digital environments that address real-world limitations.

September 11, 2025

Ryan Gehm

Gatik and Isuzu plan to produce SAE L4-capable medium-duty trucks for middle-mile applications. (Gatik)

Gatik Arena is a next-generation simulation platform designed to accelerate the development and validation of autonomous vehicle (AV) systems. Gatik, which targets autonomous middle-mile logistics, built and fine-tuned Arena in-house to meet specific operational and technical needs. Unveiled in July 2025, the platform is said to produce photorealistic, structured and controllable synthetic data that addresses the limitations of traditional real-world data collection.

Gatik Arena creates controllable, photorealistic simulation environments to train and validate its autonomous systems at scale. (Gatik)

Founded in 2017, Gatik plans to scale its freight-only – i.e., driverless – operations in 2025, and the Arena platform is central to this endeavor. Gatik collaborates with NVIDIA to integrate its Cosmos world foundation models (WFMs), which enable the creation of ultra-high-fidelity, physics-informed digital environments for robust AV training and validation, said Norm Marks, VP of Global Automotive, NVIDIA.

“NVIDIA Cosmos has been purpose-built to accelerate world model training and accelerate physical AI development for autonomous vehicles,” Marks said in a statement. “Our collaboration is helping to accelerate the commercialization of Gatik’s autonomous trucking solution at scale.”

Adam Campbell, head of safety innovation at Gatik. (Gatik)

Simulation is used to train Gatik’s autonomous systems on driving scenarios ranging from routine tasks to rare edge cases.

“It’s not just about what we see in the real world being used to create synthetic variants of those experiences. It’s also about allowing for truly, fully synthetic datasets to be developed that challenge the system in unique ways,” Adam Campbell, head of safety innovation at Gatik, said in a PAVE (Partners for Automated Vehicle Education) virtual event in August 2025. “The ability to do that is not trivial. You have to know the exposure landscape first before you can find what was not exposed.”

Campbell added that intimately knowing the middle-mile use case reduces the “landscape of unknowns.” “That’s a far cry from other use cases of autonomy,” he said. “Not to pick on robotaxis, but they have a much broader autonomy problem to solve for – not knowing where you’re starting and ending a route means that the validation of that system has to consider all possible – effectively infinite – combinations of start and end points even in a small geofenced region of a major city.” It’s not an impossible challenge to solve for, Campbell said, but the cost, scale and requirements to execute in that type of environment are expensive and time consuming.

Reducing – not replacing – on-road testing

Gatik collaborates with NVIDIA to integrate its Cosmos world foundation models. Physical AI developers use WFMs to generate custom synthetic data or downstream AI models for training robots and autonomous vehicles. (NVIDIA)

The Mountain View, California-based company aims to reduce its reliance on on-road testing to accelerate the commercialization of its autonomous solution for partners such as Kroger, Tyson Foods and Loblaw. But physical testing is not going away anytime soon.

“We’re not in a place yet or in the foreseeable future to say that all systems and safety validation work will happen in simulation. I think that’s a pipedream for many, many years ahead when even more sophisticated modeling takes place,” Campbell said. “This is real software that’s being loaded onto a real truck that is going to be moving at speed on public roads. So there is an end point where the transfer of the testing activities does have to take place on the actual physical system in the physical environment. But reducing that dependency and maintaining that high bar of safety and integrity is the goal.”

Arena is integrated with Gatik’s autonomy stack and live safety case platform, allowing teams to validate behavior, assess system robustness and close safety-assurance gaps. (Gatik)

Capturing rare events in the real world can be expensive, time-consuming and unsafe, Campbell said. Arena addresses this with high-fidelity synthetic data generation – combining real-world logs, trajectory editing, agent modeling, and multi-sensor simulation pipelines to deliver full closed-loop simulations. The result, he said, is scalable, safe and repeatable AV testing in the digital world.

“As the AV industry pushes toward scaled deployments, the bottleneck isn’t just better algorithms – it’s better, smarter data,” said Gautam Narang, Gatik’s CEO and co-founder. “Arena allows us to simulate the edge cases, rare events and high-risk scenarios that matter most, with photorealism and fidelity that match the complexities of the real world.”

Gatik is using the NVIDIA DRIVE AGX platform, accelerated by the DRIVE Thor system-on-a-chip, across its fleet of medium-duty autonomous trucks. (NVIDIA)

Arena is engineered to simulate difficult-to-collect and safety-critical scenarios with spatiotemporal consistency and sensor-level realism, Narang said. Such scenarios include:

Adverse weather and visibility – Rain, fog, snow, low-light, glare and occlusion impacting perception.
Unpredictable road users – Jaywalking pedestrians, weaving cyclists, lane-splitting motorcycles, animals and erratic drivers.
Challenging road geometry – Unprotected turns, faded markings, roundabouts and poorly marked intersections.
Dynamic road changes – Construction, detours, school zones, emergency vehicles and temporary traffic shifts.
Sensor and perception failures – Occluded signs, low-contrast objects, lidar noise, reflections and degraded sensor inputs.
Dense urban interactions – High-traffic, mixed road users, double parking and limited maneuvering space.

“Perfection is not always the goal in simulation. Sometimes what you want to do is make it less idealistic,” Campbell said. “We want to challenge the system. For example, on the lidar side, which uses laser scanning to create three-dimensional point cloud representations of the world, we’ll create highly reflective environments that make lidar-based imagery challenged – it’ll create scatter, it’ll limit the amount of returns to the system. So we train the system to operate in less-than-idealistic ways.”

Advanced AI techniques

Gatik claims that traditional fleet testing and data logging cannot provide the scale, diversity or reproducibility required to validate AV systems comprehensively. Arena addresses this shortcoming through an extensible modular simulation engine that leverages advanced AI techniques (NeRFs, 3D Gaussian splatting, diffusion models) to enable scaling from quick scenario augmentation to high-fidelity, full-stack simulation.

The company says that it’s critical to train and validate each component of the “autonomy pipeline” with the optimal technique:

Photorealistic neural rendering – Leveraging neural techniques like volumetric reconstruction to create high-fidelity simulations from abstract representations such as segmentation maps, lidar and HD maps.
Scenario editing and control – Support for modifying real-world logs – adjusting traffic flow, pedestrians, lighting and road layouts to conduct controlled A/B testing.
Sensor-accurate outputs – Multi-modal simulation across camera, lidar and radar to reflect real-world sensor behavior under varied environmental conditions.
Closed-loop simulation – Real-time interaction between ego-vehicle decisions and surrounding agents, or non-player characters (NPCs), enabling testing of the full autonomy stack in complex interactive environments. This includes modeling vehicle dynamics, policy interactions and latent scene evolution.
Structured synthetic data generation – Enables scalable, annotation-free data generation for machine learning workflows, regression testing, and safety case validation.

Arena plays a central role in Gatik’s broader development strategy, Narang said. The platform is integrated with the company’s autonomy stack and live safety case platform, allowing AV teams to validate behavior, assess system robustness and close safety-assurance gaps.

“With Arena, we’re reimagining simulation not just as a testing tool, but as a core enabler of safe, scalable autonomy,” Narang said. “It gives us the control, realism and flexibility we need to rapidly build confidence in our systems.”

Gatik, Isuzu to produce SAE L4 trucks

Gatik also announced in March 2025 that it will use the NVIDIA DRIVE AGX platform, accelerated by the DRIVE Thor system-on-a-chip, across its fleet of medium-duty SAE Level 4 autonomous trucks. The automotive-grade DRIVE AGX Thor, which runs on the safety-certified DriveOS operating system, is built on the NVIDIA Blackwell architecture. DRIVE AGX Thor delivers more than 1,000 trillion operations per second of high-performance compute, supporting the onboard AI-processing necessary to deploy Gatik’s trucks.

The collaboration with NVIDIA builds on a partnership Gatik began with Isuzu North America in 2021. Three years later Gatik and Isuzu committed to collaborating on the development and production of L4-capable autonomous trucks with safety-critical redundant systems (including braking, steering and sensors, as well as software) with complete validation required for autonomous operation.

This truck platform will reportedly be manufactured “on a first-of-its-kind production line” at a new Isuzu production facility in South Carolina. Expected to come online in 2027, the facility is anticipated to have an annual production capacity of approximately 50,000 vehicles by 2030. Isuzu says the variable-model, variable-volume production system will help usher in the next generation of autonomous, connected and carbon-neutral solutions.

Gatik’s medium-duty autonomous trucks are commercially deployed in multiple markets including Texas, Arkansas, Arizona and Ontario, Canada.

“The [middle-mile] use case allows for more direct application of that testing and validation activity,” Campbell said. “It’s always the unknown unknown things that those who wear safety hats are very much concerned about. That dimension gets narrowed exponentially because of the known route structure and the defined networks in which we deploy.”