A new arXiv preprint from researchers at Applied Intuition describes a driving simulator that maintains a throughput of 1.3 million agent-steps per second using one server-grade GPU, and employs it to train autonomous driving policies from nothing, with no human demonstrations at all [S1]. The paper, posted on 15 July 2026, reports that the resulting policies lead the InterPlan long-tail benchmark, achieve the lowest collision rates on routine-driving val14, and transfer without retraining to cities absent from the training set [S1]. The catch: none of this has been peer-reviewed, and every benchmark claim is author-reported [S1]. But if the numbers hold, the work points at a way to train self-driving AI that does not depend on logging millions of miles of human driving. How does a policy learn to drive without ever watching a human do it, and what does that mean for the companies spending billions trying to?

The data problem TerraZero sidesteps

Most autonomous driving research leans heavily on logged human driving data. Fleets collect millions of miles of real-world footage, and models learn to imitate what humans did in similar situations. The approach works until it does not: rare events, the long-tail scenarios where crashes actually happen, barely appear in the logs. You cannot train a policy to handle a situation it has almost never seen.

TerraZero takes a different route. It treats logged data as nothing more than a source of real-world map geometry, the road layout itself [S1]. Everything that moves on those roads, the other cars, the pedestrians, the cyclists, the traffic signals, is generated procedurally. For each map, the system fills in procedurally generated traffic participants and signal controllers governed by rules, while also varying the physical dynamics, reward signals, and vehicle dimensions from one episode to the next [S1]. One real map yields an unbounded set of scenarios because the traffic on it changes every time.

The training itself is pure reinforcement learning. All policies in the study are trained from nothing via self-play across GPUs, using no human demonstrations and with no hand-coded fallback planner during inference [S1]. The model learns by trying, crashing, then adjusting, not by copying a human driver.

Why 1.3 million steps per second matters

The engine behind TerraZero is a configurable C program that splits the workload: the CPU handles simulation logic while the GPU handles policy inference, with the two sharing memory directly rather than copying data between them [S1]. The result is 1.3 million agent-steps per second on one server-grade GPU, which the authors describe as far faster than existing object-level simulators [S1].

Speed in simulation is not a vanity metric. Reinforcement learning is notoriously sample-inefficient; a model might need billions of interactions to learn a complex skill. At 1.3 million steps per second, a single GPU can generate the volume of experience a policy needs in hours rather than weeks. That changes the economics of training. A two-person robotics firm that previously could not afford the compute to run enough simulated driving miles suddenly has a path that fits on one machine.

The engine also keeps fidelity features that lighter single-agent simulators skip: mixed road-user types interacting simultaneously, distinct dynamics for different vehicle categories, and consistent enforcement of traffic rules [S1]. The same system does two jobs: it produces driving policies for both cars and trucks, and it also generates sim agents capable of controlling not just vehicles but pedestrians and cyclists together [S1].

What it means

The core claim is that you do not need human driving data to train a competent driving policy. You need real maps, a fast simulator, and enough randomized scenarios. The model figures out driving through trial and error in a world where every episode throws different traffic at it.

The reported results are striking. On the InterPlan long-tail benchmark, TerraZero becomes the first entirely learned policy to claim the leading position, outperforming bigger learned planners [S1]. On routine-driving val14, it places among the top methods and records the lowest collision and time-to-collision figures, making it the safest performer in that evaluation [S1]. On Waymo Open Sim Agents realism, the approach beats other methods that also skip demonstrations and holds its own against the top method that relies on reference data [S1].

Perhaps the most curious finding: the policies transfer to untrained cities and datasets, and even develop left-hand-traffic behavior without any explicit instruction to do so [S1]. The model was never told how to drive on the left side of the road. It figured it out from the rules and the traffic patterns it encountered. That kind of unsupervised generalization is exactly what the autonomous driving industry has struggled to achieve with imitation learning, where a model trained on California right-hand traffic can fail badly in Sydney.

What it means for business

For autonomous vehicle companies, the appeal is direct. Logging fleets is expensive. A company running 500 cars with drivers in shifts burns through capital collecting data that is mostly redundant, the same highway miles, the same stop-sign turns. If TerraZero's approach holds up, the bottleneck shifts from data collection to compute. One server-grade GPU running 1.3 million steps per second is a different cost structure than a fleet of instrumented vehicles.

For simulation vendors, the bar just moved. A C engine that runs on CPU and GPU with a zero-copy path, sustains 1.3 million steps per second while enforcing traffic rules, is a serious engineering achievement. Python-based simulators that trade speed for ease of use now have a competitor that does not make that trade-off.

For smaller operators, a suburban robotics shop or a logistics firm testing last-mile delivery robots, the zero-demonstration angle matters. You do not need a team of safety drivers logging miles in your target city. You need the map geometry, which is often available from open datasets, and the compute to run self-play. A two-person team could prototype a driving policy on a single GPU server.

For insurers and regulators, the claim that a fully learned policy with no fallback planner posts the best collision scores on a standard benchmark raises a question they will need to answer: if the model is safer than alternatives that include human-designed safety layers, does the regulatory framework still require those layers?

What we don't know yet

The paper is an arXiv preprint that has not been peer-reviewed [S1]. Every performance claim, the 1.3 million steps per second, the InterPlan ranking, the collision scores, the zero-shot generalization, is author-reported. The comparative speed and fidelity claims do not name specific baselines in the provided material, so it is hard to judge the margin against existing simulators.

The emergent left-hand-traffic driving is described without external corroboration. It is an author-reported observation, and the mechanism behind it is not fully explained in the available evidence. Whether the model truly learned to drive on the left from general principles, or whether the procedural randomization happened to include left-hand traffic scenarios, is unclear.

The benchmark superiority claims, first on InterPlan, best on val14, may be contested once other researchers can run the same tests. The authors are affiliated with Applied Intuition, a company that sells simulation tools to the autonomous driving industry, which means the work sits at the intersection of research and product marketing.

What would settle the questions: an independent reproduction of the 1.3 million steps per second figure on equivalent hardware, a peer-reviewed evaluation on InterPlan and val14 by a group without ties to Applied Intuition, along with open access to the procedural generation code so others can inspect how scenarios are randomized. The next concrete event to watch is whether the code is released, and whether a competing group reproduces the zero-shot generalization claim on a city neither they nor the authors have tested before.

If you found this useful, subscribe for the next one.

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.