The Autonomous Trucking Moment
Raquel Urtasun, the former Toronto AI professor who founded autonomous trucking startup Waabi, is not given to hype. Her career spans years of foundational machine learning research, leadership of Uber's Advanced Technologies Group, and now building one of the most technically ambitious self-driving truck programs in the world. When she says Level-4 autonomous trucks are approaching commercial viability, the statement carries weight that more speculative claims in the autonomous vehicle space have not.
In an extended interview with IEEE Spectrum, Urtasun outlined Waabi's technical approach, its progress toward commercial deployment, and her view of how generative AI has fundamentally changed the timeline for achieving the kind of robust, generalizable autonomy that makes long-haul trucking a viable application for full self-driving technology. Her argument is not that the problem has become easy, but that the tools available to solve it have improved dramatically.
Level-4 autonomy — the capability to handle all driving tasks within a defined operational design domain without any human intervention — is the threshold that separates demonstration technology from commercial product. For trucking applications, the relevant domain is primarily highway driving on defined routes, a substantially more constrained environment than the complex urban settings that have challenged passenger vehicle autonomy programs for years.
The Generative AI Advantage
Urtasun's core argument is that generative AI approaches to self-driving — which use large models trained on vast amounts of driving data to learn generalizable driving behaviors rather than encoding explicit rules — have produced qualitative improvements in the robustness of autonomous systems in ways that previous approaches struggled to achieve. The same scaling dynamics that produced GPT-4 and its successors are now being applied to the driving problem, with comparable step-change results in capability.
Waabi's architecture centers on what the company calls a generative world model — a learned simulation environment that can generate realistic driving scenarios including rare and dangerous edge cases that would be too dangerous or expensive to encounter and record in real-world data collection. This simulation capability addresses one of the most fundamental bottlenecks in autonomous vehicle development: the need for training data covering the full distribution of situations the system might encounter in deployment, including low-probability events that have disproportionate safety implications.
The ability to use learned simulation to stress-test autonomous systems against a virtually unlimited variety of generated scenarios means that Waabi and similarly-architected programs can cover far more of the tail risk distribution than programs dependent on recorded real-world data. For safety certification and regulatory approval, this is not merely a development efficiency advantage — it is a fundamentally different approach to demonstrating that a system is ready for deployment.
