In 2005, Drs. Katalin Karikó and Drew Weissman published a paper showing that chemically modified messenger RNA could evade the body’s innate immune sensors, avoiding the dangerous inflammatory response that made synthetic mRNA useless as a therapeutic. Nature rejected the manuscript without review. So did Science. It took 15 years, hundreds of researchers, and the maturation of lipid nanoparticle delivery and scalable manufacturing before that finding could underpin a vaccine.
When COVID-19 arrived, a vaccine was designed to detect candidate sequence in days, but only because decades of foundational work had already been done.
This recurring pattern is common across transformative technologies. Initially, a concept is shown to be feasible in limited settings. Then comes a long stretch where crucial tools lag, stalling progress. During these phases, funding wanes and skepticism grows. Finally, when tools from related fields advance, these stalled concepts can suddenly move forward and become real.
Autonomous trucking is among the most important applications of physical AI: trucks that drive themselves on real highways carrying real freight. From the outside, this sector looks stalled. It is not. Its technical foundations have been quietly transformed, largely unnoticed by the investors and analysts who once championed it.
13 runs
Before the current wave of commercial deployments, the entire industry had produced roughly 13 driverless runs on public roads. Every company combined.
Ten were empty-cabin runs conducted that ran on Arizona desert routes during nighttime hours when traffic was minimal and conditions were most favorable. These proved that a truck could navigate a highway without a human at the wheel, but they were demonstrations of feasibility under the most constrained circumstances. The remaining three retained a safety operator in the cabin.
Thirteen runs. That was the state of the art at the time. The natural question is: if this works now, why did it take so long?
Not a bigger robotaxi
From the outside, autonomous trucking looks like the robotaxi problem with a bigger vehicle. It is a fundamentally different problem.
A fully loaded Class 8 truck weighs 80,000 pounds. Its stopping distance at highway speed exceeds 500 feet, about one and a half football fields. A truck must therefore perceive and plan over a long distance, usually 550 yards, approximately triple that of a passenger vehicle, and this single requirement has cascading consequences. At this range, camera-based depth estimation degrades with distance due to geometric constraints, and a single pixel error can translate into an error of 30 yards. Lidar sensors capable of the necessary range, resolution, and refresh rate at highway speed did not exist in a commercially viable form until recently. Radar provides range but lacks spatial resolution for scene understanding at a distance. Each modality hits theoretical ceilings at trucking-relevant ranges, and fusing them at extended distances introduces compounding uncertainty that must be handled in the network architecture itself. For much of the past decade, the on-vehicle GPU hardware could not accommodate networks of the required size at the required latency.
The research community compounded this gap. The major public datasets, benchmark competitions, published architectures, and evaluation metrics are almost entirely oriented around urban robotaxi driving at short to medium range. Many core technical problems in autonomous trucking, particularly long-range perception beyond 200 meters, multi-agent prediction at highway speed, and heavy-vehicle dynamics under wind and load variation, have never been systematically addressed in the academic literature. A company serious about this problem cannot adopt off-the-shelf research; it must build significant portions of the stack from first principles.
Then there is the operational envelope. A robotaxi covers mapped city streets. By contrast, a truck traverses thousands of highway miles, navigating construction zones, changing road surfaces, and varying traffic at high speeds. Rain is an occasional nuisance for a robotaxi. A truck driving eight hours will almost certainly encounter multiple weather transitions, including abrupt sensor degradation when entering a rainstorm at 65 miles per hour with every camera and lidar surface covered in spray.
What has changed
Earlier entrants attempted autonomous trucking with AI systems built on human-labeled driving data and neural network models with limited capacity. These systems worked for the scenarios they had been explicitly trained on, but could not reliably handle novel situations. It is a bit like attempting transoceanic aviation before the jet engine: the route map was correct, but the propulsion technology had not arrived.
Every technical limitation converged on a single operational reality: earlier systems could only function safely on the easiest roads, in the best conditions, at the quietest hours. The technology constrained the operation to a narrow envelope, and within that envelope, there was no viable commercial business.
What has changed is the propulsion technology, and the change is recent. Three shifts have converged since roughly 2022.
Foundation models and world models. The previous generation of autonomous driving perception was built on task-specific convolutional neural networks trained from scratch on narrow driving datasets. The shift to transformer-based architectures, combined with large-scale visual pre-training, has fundamentally changed what is possible. Contrastive Language-Image Pre-training (CLIP) model and its successors demonstrated that visual representations trained on diverse internet-scale data transfer powerfully to downstream tasks, including those never seen during training. Segment Anything Model (SAM), trained on over 1 billion masks across 11 million images, showed that a single visual foundation model could generalize zero-shot to entirely new image distributions and visual domains. These advances matter for trucking because the core perception challenge—reliably understanding a driving scene across weather, lighting, and road conditions the system has never encountered—is precisely the kind of generalization problem that pre-trained visual representations now solve. Modern autonomous driving perception stacks leverage these pre-trained backbones as feature extractors, then fine-tune with driving-specific heads for tasks such as 3D detection, occupancy prediction, and lane estimation at both short and long range. Equally important, world models trained on top of these representations predict the behavior of other road agents, such as trucks, cars, merging vehicles, erratic drivers, enabling reliable highway-speed planning over long horizons.
Large-scale reinforcement learning. The same paradigm has arrived in physical AI, but the key enabler was scale. Earlier attempts at RL for driving were limited by sample efficiency: the algorithms required billions of environment interactions to converge on robust policies, and generating those interactions was computationally infeasible. What changed was the arrival of massively parallel GPU simulation. Platforms running on certain hardware now allow thousands of simulation environments to run concurrently on a single GPU cluster, generating on the order of 10 billion samples in a day. This is the regime in which on-policy methods such as Proximal Policy Optimization (PPO) become viable for complex decision-making and physical control. The approach has already been validated in adjacent domains: in humanoid robotics, transformer-based controllers trained with large-scale PPO across thousands of randomized simulation environments have been deployed to real-world full-sized humanoid robots zero-shot. The principle transfers directly to trucking: a driving policy trained via super large-scale RL in simulation can encounter and master the rarest, most dangerous scenarios, such as a tire blowout in an adjacent lane, a highway closure, a vehicle cutting across three lanes, millions of times before encountering them on the road. The limiting factor is no longer how many miles a fleet can accumulate but how much compute can be directed at the problem, and compute scales.
The hardware ecosystem. Lidar sensors have reached the range, resolution, and refresh rate that trucking demands. Current-generation units deliver 300-meter detection range at 10% reflectivity with 350 channel resolution. Five years ago, neither the sensors nor the silicon existed at the performance, reliability, and cost required for commercial-scale deployment. Today, Tier 1 and Tier 2 automotive suppliers have matured their autonomous-vehicle component offerings, and the supply chain is ready in a way it simply was not before.
These three shifts are what separate the current generation of autonomous trucking systems from their predecessors. Companies building on these foundations from Day 1 can train in simulation at a density no real-world-only fleet could accumulate, perceive at the distances trucking demands, and run the resulting models on hardware that fits inside a truck.
Why it matters—and the road ahead
Trucking moves 72% of all freight tonnage in the United States. Its costs are embedded in the price of groceries, building materials, and medicine. The industry faces a projected shortage of 160,000 safe drivers by 2030, a structural deficit already pushing up freight rates and consumer prices. At nearly $1 trillion in annual U.S. revenue, with surging but unmet demand, a technology that addresses the binding constraint in this industry is one of the largest economic opportunities in the industrial economy.
The tools have arrived. The economic need is acute. What remains is execution, and execution, unlike fundamental research, requires focus, resources, and time.
