TAMPA, Fla. (BLOOM) – Before the next generation of self-driving cars hits the road, it’s already logging millions of miles, virtually. Inside labs and data centers, autonomous vehicles are learning how to navigate sudden snow squalls, jaywalking toddlers, malfunctioning traffic lights, and other rare edge cases that would be too costly, or too dangerous, to recreate on public roads. It’s just another day at the (simulated) office.

One of the experts making these virtual environments possible is Zhuochun Liu. At NVIDIA, she develops autonomous vehicle simulation platforms like DRIVE Sim, using their Omniverse technology. These tools allow automakers and AI developers to design, test, and validate the next generation of autonomous systems. “If we’re talking about what it takes to adopt and scale autonomous vehicles,” she says, “simulation is the safest, most efficient place to start.”

What Is a Digital Twin, and Why Do AVs Need One?

A digital twin is a high-fidelity virtual version of a real-world system, designed not just to look the part, but to behave like it. As opposed to typical 3D models used in gaming or design, digital twins incorporate accurate physics and system dynamics. For AVs, that fidelity is important, capturing everything from tire friction on wet pavement to the latency in radar signal processing.

Digital twins allow developers to recreate and analyze obscure but high-risk scenarios—say,  a road cyclist emerging from behind the mountain pass, or glare off a rain-slick city street—and evaluate how perception algorithms and control systems respond. More than just rerunning footage, they can tweak environmental variables, modify sensor placements, and isolate individual system reactions, down to the millisecond lag between neural net inference and actuator output.

“With hardware, every new configuration means rebuilding and revalidating the physical vehicle,” Liu explains. “We’re trying to displace all of that into a digital sandbox, where your iterations don’t carry that material cost.”

A Safer Way to Scale

Automakers are all in. A 2023 survey found that three-quarters of automotive organizations were already using digital twins in their R&D workflows. While physical testing still plays a role, it’s narrower in scope and increasingly focused on confirming what’s already been validated in simulation.

A single day in simulation can compress what would otherwise take months of real-world testing. Developers can recreate historical failures and evaluate dangerous situations that might otherwise take millions of real-world miles to encounter. This is especially important in the wake of newly proposed NHTSA rules that would require AV systems to meet higher crash standards. Platforms like DRIVE Sim let companies rehearse complex behaviors in controlled environments, reducing both the cost and risk of live trials.

Since mid-2024, Liu has led efforts to advance NVIDIA’s autonomous vehicle platform through simulation, turning the company’s sim-in-the-loop system into a production-grade tool for automakers. Her work extended the pipeline’s coverage to include L2 and L2++ features—features like cruise control and lane centering which serve as critical steps toward full automation. Today, these integrations are a core part of how the platform accelerates product delivery and reduces reliance on costly in-vehicle testing for external partners.

Automakers Are All In

This shift isn’t confined to tech-forward Silicon Valley startups. Legacy automakers have already restructured many of their engineering stacks around simulation-first development.

Jaguar Land Rover, for instance, is building its 2025 lineup on a software-defined architecture co-developed with NVIDIA, with all testing validated in simulation long before physical prototypes roll out. Mercedes-Benz, meanwhile, uses simulation to support software development and system stability, running thousands of virtual driving scenarios to catch bugs and fine-tune behavior before it reaches production vehicles. These digital twins, based on NVIDIA Omniverse, allow automakers to deliver safer, more reliable systems while minimizing delays and post-production risks.

These simulation-first strategies are especially valuable in AV programs targeting Level 3 or Level 4 automation, where the car handles full situational awareness, with all its nuances. “To validate those systems, you need billions of data points across millions of scenarios,” Liu says. “Simulation lets us compress that timeline while building confidence in how systems behave under stress.”

The shift to simulation has also been accelerated by necessity. Supply chain disruptions in recent years have slowed hardware production and limited vehicle availability for testing. Meanwhile, the complexity of automotive software has ballooned, with McKinsey estimating it’s been growing by 40% per year since 2021. With most modern cars defined by code, virtual testing environments have become central to how that code is built and validated.

Even U.S. policy is beginning to reflect this reality. While the CHIPS and Science Act is best known for supporting semiconductor manufacturing, it also directs funding toward simulation research and digital twin infrastructure—part of a broader push to bolster domestic competitiveness in advanced manufacturing and mobility.

Where We’re Headed Next

NVIDIA’s automotive segment reported record quarterly revenue of $570 million in Q4 FY25, evidence of how seriously automakers are investing in simulation-backed platforms. Liu likens it to what happened in aviation. “No one would fly a plane today that wasn’t first tested in simulators, by both the engineers and the pilots. Cars are headed in that direction. We’re just adapting to the complexity that autonomy demands.”

Where simulation once served to fine-tune a nearly finished system, it’s now often the starting point. Thanks to digital twins and virtual testbeds, developers can stress-test edge case AI reactions and rehearse over-the-air software updates. Virtual environments are, in turn, shaping how the real systems behave, where the learning isn’t based on failures and trial-and-error and readiness isn’t left to chance.

“Simulation is so integral that the terminology is beginning to flip,” Liu notes. “It’s not uncommon for automotive partners to talk about ‘hardware-in-the-loop.’”

Autonomy remains a long-term goal, but the most important leg of the journey toward autonomy, from edge case discovery to sensor tuning to regulatory validation, is already taking place in a world that runs on code over concrete.