Waymo's lidar stack, Tesla's vision-only bet, and Zoox's wheel-free pod are chasing the same goal by different routes — while the AI quietly retiming traffic signals and coaching truck drivers is already running at far greater scale.
In the summer of 2025, Austin became the first US city where you could hail a ride from two different companies' fully driverless cars and get a materially different experience depending on which app you opened. Waymo's Jaguar I-PACEs, bristling with roof-mounted lidar, drive with a safety operator nowhere in the car. Tesla's Model Y robotaxis, running on cameras alone, launched with a human safety monitor in the passenger seat — a concession regulators and skeptics both read as a tell. By early 2026, Amazon's Zoox had added its purpose-built, steering-wheel-free pod to the mix in Las Vegas. Three companies, three sensor philosophies, three different bets on what "autonomous" is allowed to mean before the public will get in the car.
Transportation is one of the few verticals where AI's promises and its failures are both extremely visible — a stuck robotaxi blocking an intersection makes the news in a way a bad customer-service chatbot never will. That visibility is useful: it forces honesty about where the technology actually works today versus where it's still a demo with a safety driver. Four application areas show the real state of play: urban robotaxis, highway trucking, traffic signal optimization, and fleet telematics. They are not equally mature, and the gap between them is the story.
The core technical split in urban autonomy is whether you build a 3D model of the world directly from sensors (lidar + radar + cameras) or infer it purely from camera video, the way a human does.
Waymo runs a sensor-fusion stack: lidar for precise depth and object boundaries, radar for velocity in bad weather, and cameras for classification (reading signs, traffic lights, pedestrian intent). This is expensive — lidar units alone used to run into the tens of thousands of dollars per vehicle, though costs have fallen sharply as the company moved to its own in-house lidar design. The payoff is redundancy: if cameras are blinded by glare, lidar still sees. Waymo operates commercial driverless service (no safety driver) in Phoenix, San Francisco, Los Angeles, and Austin, and has been expanding into freeway driving and additional metros, after years of accumulating real-world miles in a tightly geofenced, HD-mapped operational design domain (ODD).
Tesla bets that cameras plus a large enough neural network can match or beat lidar-based perception, at a fraction of the hardware cost, and that this is the only approach that scales to tens of millions of consumer vehicles rather than a fleet of purpose-built taxis. Tesla's Robotaxi service launched in Austin in June 2025 on this vision-only stack, initially with an employee safety monitor in the passenger seat and a geofenced service area — a materially more cautious rollout than the "button-press, no human anywhere" framing in Tesla's marketing. The company has faced NHTSA inquiries into Full Self-Driving behavior in prior years, and the honest read is that vision-only has made real progress but has not yet demonstrated the same depth of unsupervised, no-human-in-the-loop track record as Waymo's lidar stack.
Zoox (Amazon) skips the retrofit question entirely: its vehicle has no steering wheel or pedals, bidirectional symmetric design, and four-corner sensor pods for 270-degree coverage with no blind spot behind the car. It's the most radical hardware bet and the least proven at scale — service so far is limited to small zones in Las Vegas and San Francisco.
The practical takeaway: lidar-based fusion is currently the more battle-tested approach for true driver-out operation, vision-only is the lower-cost long-term bet still proving itself, and purpose-built robotaxi hardware is the furthest from scale. All three remain geofenced — none operates "anywhere" the way a human driver does, and weather (heavy rain, snow), construction zones, and unmapped roads remain the hard edges of every ODD.
If urban robotaxis are the hard case, long-haul highway trucking is the version of autonomous driving that's closer to solved — and it's quietly further along in actual driver-out operation than most consumers realize.
Aurora Innovation began driverless (no one in the cab) commercial trucking on the Dallas–Houston corridor on Texas highways in 2025, after years of supervised pilot hauls with major freight customers. Kodiak Robotics has pursued a similar highway-focused strategy, including a defense/off-road variant. The reason trucking moves faster than robotaxis isn't that trucks are easier to control — it's that the operational design domain is dramatically simpler: limited-access highways have no pedestrians, no cross-traffic, no double-parked delivery vans, and comparatively predictable merge geometry. The hard part of urban driving — social negotiation with other road users — mostly doesn't exist at 70 mph on I-45.
The current model for most of these companies is "highway autonomy, human-driven first/last mile": the truck drives itself between transfer hubs near the highway, and a human (or a separate local carrier) handles the surface-street leg into the depot. That hybrid structure is a limitation, not a bug — it's the honest boundary of where the technology is actually ready, and it maps directly onto the geofencing problem robotaxis face, just with a much friendlier ODD.
While robotaxis get headlines, the AI application with the broadest quiet footprint in transportation might be Google's Project Green Light, which uses aggregated, anonymized Google Maps driving data and machine learning to recommend retimed traffic signal plans to city traffic engineers — without requiring new hardware at the intersection. The model looks at how traffic actually queues and clears at a given light across many trips, flags where signal timing causes unnecessary stop-and-go (a major source of both delay and emissions), and proposes adjusted phase timing that a city engineer reviews and can push live. It has been deployed across a growing number of cities internationally since its original pilot, with Google reporting meaningful reductions in stopping and idling at optimized intersections — real but modest, incremental gains, not a reinvention of traffic engineering.
This is a useful contrast case: it's AI applied to a narrow, well-bounded optimization problem (signal phase timing) using data the company already collects for an unrelated product (Maps), deployed with a human engineer firmly in the loop and the ability to revert. It's unglamorous and it's actually running in production in far more places than any robotaxi fleet.
The least visible but most economically pervasive transportation AI is in commercial fleet management. Companies like Samsara and Motive sell AI dashcams that run computer vision in the cab to detect distracted driving, following distance, and drowsiness in real time, coaching drivers and feeding safety scores back to fleet managers — and increasingly to insurers, who use those scores to adjust commercial auto premiums. Separately, predictive-maintenance models trained on vehicle telematics (engine sensor data, brake wear patterns, tire pressure trends) flag likely failures before they cause a breakdown or accident, which matters enormously for fleets where downtime is the dominant cost.
This is AI in transportation with none of the controversy of robotaxis, because there's still a human driver and the AI's job is narrowly coaching and prediction, not control. It's also the category generating the most immediate, measurable ROI for the companies buying it — the business case is accident reduction and maintenance cost, not a futuristic mobility vision.
| Application | Technical approach | Scale today | Primary limitation |
|---|---|---|---|
| Urban robotaxis (Waymo) | Lidar + radar + camera fusion | Commercial, driver-out, multiple US metros | Geofenced; weather and construction edge cases |
| Urban robotaxis (Tesla) | Vision-only neural network | Early commercial, safety monitor present | Shorter unsupervised track record than lidar stacks |
| Purpose-built robotaxi (Zoox) | Custom vehicle, 360° sensor pods | Pilot zones only | Smallest service footprint, newest to market |
| Highway driverless trucking | Lidar/camera fusion, highway-only ODD | Driver-out on select corridors | Requires human-driven first/last mile |
| Traffic signal optimization | ML on aggregated navigation data | Deployed across many cities globally | Needs city engineer buy-in; incremental gains |
| Fleet telematics / dashcams | In-cab computer vision, predictive maintenance | Broad commercial deployment | Driver assistance only, not control |
The pattern across all four: AI in transportation succeeds fastest and most durably where the operational domain is simplest and a human stays in or near the loop, and it's still working through the hardest, most consequential case — fully driver-out operation on unconstrained city streets — one geofenced metro at a time.
If you're evaluating a transportation-AI claim, ask two questions before anything else: what's the actual operational design domain (a specific geofenced area and set of conditions, or "anywhere"?), and is a human still in or near the control loop? Robotaxi and driverless-trucking announcements that answer both questions honestly are making real, measurable progress; ones that gloss over the ODD are usually still at the demo stage, regulatory approval or not.