Generative AI redefines what it means for machines to think, decide, and act in the physical world. As autonomous vehicles become a proving ground for this technology, success will depend on systems that can reason, adapt, and operate with measurable safety at scale.
Machines that can think are no longer confined to virtual environments. As the boundary between code and concrete dissolves, a new era is unfolding, where generative AI steps into the physical world, with implications as vast as the logistics networks it seeks to transform.
Self-driving is not a moonshot; it is the next market shift. When Raquel Urtasun, Founder and CEO of Waabi, took the stage at NVIDIA GTC 2025, her vision was clear: the next frontier for AI lies not in chatbots or digital assistants but in physical machines that move through the world with intelligence, speed, and rigour. The implications stretch far beyond novelty. This is not an experiment in autonomy; it is a redefinition of infrastructure, logistics, and industrial scale.
Nowhere is this ambition more tangible than in long-haul trucking. It is, Urtasun explains, where economic urgency meets technological viability. “It is a trillion-dollar market in North America alone,” she says. “The industry is under enormous pressure. Carriers face chronic driver shortages, with turnover rates reaching 90 per cent annually. Safety concerns are high, inefficiencies are rampant, 20 to 30 per cent of truck miles are driven empty. AI can save lives, reduce waste, and solve a labour crisis.”
Unlike passenger vehicles, which demand hyper-local navigation and contend with complex urban behaviour, long-haul trucking operates within more predictable constraints: highways, waypoints, and logistics hubs. This makes it an ideal proving ground for high-capacity AI systems judged not by novelty but by uptime, fuel use, and asset efficiency.
“Today, a human-driven truck might clock 80,000 miles annually, constrained by service hours and human needs,” Urtasun explains. “An autonomous truck can run nearly continuously. That translates into three or four times the asset utilisation, with additional gains in fuel efficiency, insurance savings, and route optimisation. For carriers, the appeal is commercial, not conceptual.”
Old maps, new roads
Yet, solving the trucking challenge has demanded a break from legacy thinking. Much of the autonomous vehicle industry has relied on what Urtasun calls AV 1.0, highly engineered stacks of disconnected AI modules bolted together like a pile of spaghetti. Every new edge case leads to a new submodule. It is fragile, inefficient, and resource intensive.
This modular complexity creates more problems than it solves. The system becomes unscalable, requiring thousands of engineers and years of road testing. It also lacks the necessary reasoning ability to generalise in new situations, which is essential for vehicles that operate without a human fallback.
“AV 1.0 cannot get us to real autonomy,” Urtasun continues. “We need end-to-end systems that reason about the world, not just imitate human behaviour. Black-box neural networks trained to mimic human drivers are opaque, inefficient, and hard to validate. Our approach is entirely different.”
Waabi has designed what Urtasun calls a reasoning-first AI system. It does not just see the world; it interprets it, anticipates possible futures, and chooses optimal manoeuvres based on deep situational awareness. This approach reflects how people drive but with far greater consistency and computational power. It also enables transparency and explainability, which are critical requirements in safety-critical applications.
Simulated worlds for real-world safety
Training such systems cannot rely on real-world experience alone. Creating edge-case scenarios on actual roads would be impractical and ethically indefensible. Simulation becomes the cornerstone of development, but only if it goes beyond mere visual fidelity. For Waabi, that meant building a closed-loop, interactive simulation environment: Waabi World.
“Waabi World allows our autonomy system to interact with a dynamic virtual world as if it were real,” Urtasun says. “It sees, reasons, and acts, and the world responds. We simulate everything: roads, vehicles, pedestrians, weather, sensor physics, and internal system delays. It is visually and behaviourally indistinguishable from reality.”
Unlike conventional simulators that replay recorded data or test isolated components, Waabi World operates at the system level. The autonomy stack engages with the environment continuously, allowing engineers to observe behaviour, assess decision-making, and iterate in real time. However, its most powerful feature lies in the generation of test scenarios.
“We can create counterfactuals, clone real-world miles and introduce variations such as aggressive lane changes, different weather, and altered traffic density,” Urtasun explains. “We use adversarial optimisation to find the scenarios most likely to challenge the AI. That feedback loop helps us improve system robustness without brute-force testing.”
The scale of this synthetic testing far exceeds what is possible on public roads. More importantly, it generates the statistically significant datasets required to build a credible safety case. Human-driven trucks might go ten million miles without a fatality. Reaching equivalent assurance levels through physical testing alone would be logistically and financially impossible.
A new stack for a new era
The system’s capability hinges not just on simulation but on compute infrastructure. Waabi’s partnership with NVIDIA is pivotal here. Their hardware supports the high-fidelity cloud-based simulation needed to train the AI and the real-time, onboard inference that keeps a truck safe on the road. “Efficiency is a two-part problem,” Urtasun says. “First, you need the AI to operate with minimal compute overhead onboard. Second, it must learn with fewer examples. We build systems that generalise from limited data, more like humans do. That makes them vastly more efficient and more sustainable.”
Environmental responsibility is not incidental. With concerns over AI’s energy footprint growing, especially in the cloud, physical AI cannot afford to be wasteful. Edge compute, tight software integration, and energy-aware design are all essential. These systems will be deployed at scale, not in research labs.
As deployment nears, safety becomes even more central. Urtasun is emphatic on this point: “Being safe is not enough,” she says. “You must be able to prove it scientifically before you remove the human driver. That proof must stand up to scrutiny from regulators, customers, and the public.”
That is why the combination of reasoning-based autonomy and next-generation simulation is critical. Together, they create not just a safer vehicle but a testable one. A system that can demonstrate its capabilities in a measurable, repeatable way.
From retrofits to purpose-built autonomy
Yet intelligence alone is not sufficient. Hardware must evolve in parallel. Retrofitting commercial trucks with sensors and software can support early development, but it will not scale. Redundancy, reliability, and manufacturability must be baked in at the OEM level.
This is why Waabi has partnered with Volvo to co-develop the VNL Autonomous, a purpose-built, factory-integrated platform for driverless operation. Every system, from steering to braking, is designed with autonomy in mind, including redundant pathways and fail-safe mechanisms. According to Urtasun, it is the foundation for safe and scalable commercial deployment. “Without vertical integration, you cannot achieve the consistency, reliability, and cost structure required for mass adoption,” she continues. “Retrofitting works for a demo, not for a fleet. This platform is a milestone.”
As with the autonomy stack, mapping plays a critical role. High-definition maps are often criticised for their complexity and cost, but Waabi treats them as an internal sensor, a mechanism to enhance foresight and decision-making. Their proprietary mapping system can generate these maps for under $50 per mile, making the approach viable at scale.
The road ahead
Together, the pieces form a compelling narrative: a reasoning-first AI system trained in a reactive simulation environment, operating on a purpose-built platform, supported by scalable infrastructure and scientific safety validation. It is a systemic rethink, not an incremental improvement.
It also sets a precedent. The lessons from long-haul autonomy will resonate as other industries consider bringing AI into the physical world, whether through drones, warehouse robotics, or industrial machinery. Generalisation, simulation, integration, explainability: these are not features; they are foundations.
For senior executives exploring the frontier of AI, Waabi’s approach reminds them that success depends not on novelty but on rigour. The physical world is unforgiving. Systems must perform, but they must also explain, adapt, and prove themselves with evidence. Autonomy is not just about movement; it is about trust.
As AI moves from the data center into the real world, it will demand a new kind of leadership that understands the difference between artificial intelligence and artificial confidence. More than anything that will separate the serious players from the speculative ones. In the process, the very fabric of the built world, roads, fleets, workflows, and infrastructure will reshape itself around the systems that learn to drive it.
