Should you build or buy Autonomous Vehicle Simulation?
Autonomous vehicle simulation software models a vehicle's sensors, dynamics, and driving environment in a virtual world so engineers can test perception and planning software against millions of scenarios without a physical car. AV, robotaxi, and ADAS programs buy or build it to validate software before it ever touches a public road.
Should we build our own AV simulator like Waymo did, or buy Applied Intuition or dSPACE? For almost every program the right split is to buy the scenario, coverage and qualification backbone and build the vehicle-specific models, ODD scenarios and reconstruction on top. Open tools cover R&D simulation well, but the named end-to-end self-builds belong to frontier-scale teams, and the safety-case tooling is where the vendor earns its fee.
Build it, buy it, or bridge?
When building makes sense
Building your own simulation stack is a real, proven path, but only at a scale few programs reach. Waymo says it has driven more than 20 billion miles in simulation to identify the hardest situations its vehicles will face on public roads, and treats that simulator as core to its safety case rather than a supporting tool. That kind of build works when a program has its own fleet, its own sensor suite, and years of runway to develop and validate a safety-qualified toolchain internally. For a mainstream ADAS or robotaxi program without that scale, open components like CARLA and esmini cover the development and research simulation well, letting engineers iterate on perception and planning quickly. What they don't give you is the coverage-driven validation and safety-case documentation a regulator will actually accept, which is where even well-funded programs still lean on a qualified vendor backbone.
When buying makes sense
Buying the qualified backbone is the right call for almost every program that isn't operating at Waymo's scale. Applied Intuition, dSPACE, Foretellix, and NVIDIA DRIVE Sim have built the coverage-driven scenario generation, safety-qualified toolchains, and hardware-in-the-loop integration that a regulator or safety auditor actually wants to see, and reproducing that documentation trail from an open simulator is a different project than getting a car to drive convincingly in a virtual city. CARLA, which its creators built specifically as an open urban driving simulator for research, is excellent for the development work that happens before a safety case matters. The vendor backbone is what turns simulated miles into a defensible validation record once a program is ready to certify software for real roads, which is why even teams with strong internal simulation work still buy the qualification layer rather than build it.
The desk read
Build-versus-buy analysis for Autonomous Vehicle Simulation is being written. In the meantime, the framework that drives every B4 call is on the B4 Index page.
Frequently asked
Should we build our own AV simulator like Waymo did?
For almost no program, no. Waymo's simulation stack is the product of years of dedicated engineering tied to its own fleet and safety case, a scale only a handful of companies operate at. Most ADAS and robotaxi programs are better served buying a qualified backbone and building only the vehicle-specific models on top.
Is CARLA good enough for a production autonomous vehicle program?
For research and early development, yes; CARLA is a capable open-source simulator built specifically for autonomous driving research. What it doesn't give you is the coverage-driven validation and safety-qualification documentation a regulator expects before software goes on public roads, which is where a qualified vendor platform still matters.
What does 'safety-qualified' mean for a simulation toolchain?
It means the simulator's scenario coverage, sensor models, and validation process have been documented and evaluated well enough that a regulator or auditor will accept simulated testing as evidence toward a safety case. Open research tools generally lack this documentation trail, which vendors like dSPACE and Foretellix build as part of their product.
Can synthetic or AI-generated scenarios replace real-world test miles?
They reduce how many real miles you need, not eliminate them. Vendors are adding AI-driven scenario generation on top of qualified simulation cores specifically to expand coverage of rare edge cases, but regulators still expect a mix of simulated and real-world validation before approving software for public roads.