It was supposed to be the future, and it kept arriving late. Elon Musk told an audience that Tesla would have one million robotaxis on the road by the end of 2020. That milestone never came. He was not alone in his optimism. Over the preceding decade, automotive executives, Silicon Valley founders, and venture capitalists had collectively painted a picture of autonomous vehicles cascading into cities by the early 2020s, remaking urban transport and eliminating the human driver. The picture was compelling. The engineering, it turned out, was harder.
What is striking about the robotaxi story today is not that it failed. It is that it partially succeeded, in ways that are genuinely impressive and genuinely limited at the same time. Fully driverless commercial services exist. Passengers are riding in them. And yet the technology remains geographically confined, operationally constrained, and perpetually five years from everywhere. Understanding why requires looking inside how these systems actually work.
What “Autonomous” Actually Means in Practice
The language around self-driving cars has always been slippery. The SAE International framework for vehicle automation, which became a de facto industry standard, defines six levels of automation from zero (no automation) to five (full automation under all conditions). Almost every robotaxi operating commercially today functions at what the framework calls Level 4: the vehicle can drive itself, but only within a defined geographic area, known in industry parlance as an operational design domain, or ODD.
An ODD is essentially a set of boundaries and conditions the system has been trained and validated for. It might cover specific city blocks, specific weather conditions, specific speed limits. Step outside those parameters and the system either asks for human intervention or, in a fully driverless deployment, stops the vehicle and calls a remote operator. This is not a failure mode. It is the design.
Waymo, the Alphabet-owned company that has operated commercial driverless services in cities including San Francisco, Phoenix, and Los Angeles, builds extraordinarily detailed maps of its operating zones. Its vehicles carry lidar, radar, and camera arrays that feed a sensor fusion system, cross-referencing real-time perception with pre-mapped environments to localize the car to within centimeters. When a stretch of road changes, because of construction, flooding, or an unusual event, the system can struggle. Waymo addresses this with a combination of rapid map updates and remote human operators who can intervene or guide vehicles through unexpected situations.
The hardware investment alone is staggering. Early-generation autonomous test vehicles carried lidar units that cost tens of thousands of dollars each. Costs have fallen substantially over time, but a fully equipped robotaxi still carries sensor suites, computing hardware, and redundant safety systems that add significant expense per vehicle, well beyond what a standard car requires.
The “Long Tail” Problem That Humbles Every Timeline
Engineers who work in autonomous vehicles often refer to the long tail of edge cases, the enormous catalog of unusual, ambiguous, or unexpected situations that a vehicle must handle correctly, every single time, to be considered safe for unsupervised public operation.
A human driver processes edge cases through a lifetime of learned context. A child chasing a ball into the street, a construction worker waving traffic through a red light, a funeral procession, a mattress falling off a truck at highway speed: people handle these because they understand the social and physical world around them, not merely the rules of the road. Teaching a machine to do the same requires either encountering and labeling millions of such scenarios in training data, or building systems general enough to reason about novel situations without having seen them before. Neither approach is solved.
This is the core reason deadlines slip. The first ninety percent of the problem, handling routine driving in favorable conditions, can be solved with enough data and computing power. The remaining fraction, the scenarios that are rare individually but collectively happen all the time across a large fleet, is where progress becomes exponential in difficulty. Waymo has reported driving large numbers of miles in simulation, supplementing far smaller real-world datasets, precisely because gathering enough real-world edge case data would take decades at any plausible fleet size.
Cruise, the General Motors subsidiary that operated a competing driverless service in San Francisco before GM stopped funding the robotaxi business in December 2024, illustrates what happens when that long tail bites. In October 2023, a Cruise vehicle was involved in an incident with a pedestrian who had already been struck by another car. The Cruise vehicle failed to stop appropriately. The California Department of Motor Vehicles subsequently suspended Cruise’s permit to operate driverless vehicles in the state. The episode became a significant setback not only for Cruise but for industry confidence more broadly, and it accelerated scrutiny of how companies report safety incidents to regulators.
The Regulatory Patchwork That Shapes the Map
Where robotaxis can operate is almost as much a regulatory question as a technical one. The United States has no single federal framework governing autonomous vehicle deployment. Instead, authority is fragmented between federal safety standards, which historically focused on occupant protection in human-driven vehicles, and a state-by-state permitting regime that emerged partly by default.
California, home to the largest concentration of AV testing activity, developed one of the more detailed regulatory frameworks through its Department of Motor Vehicles. Permits are required for testing with a safety driver, for testing without one, and for commercial deployment. Companies must report disengagements, the moments when a human operator or system override takes control, and those reports have historically given researchers a rough window into how often systems fail to handle situations autonomously.
Arizona became a favored deployment location for Waymo partly because of its comparatively permissive regulatory environment and its dry, clear weather, which suits current sensor technology. Phoenix does not have the dense urban complexity of San Francisco, and Waymo operated there for years before expanding to more challenging environments. That sequencing was not accidental.
Texas, Florida, and several other states have passed legislation aimed at attracting autonomous vehicle operations, generally by limiting local government ability to impose restrictions. The result is a geography of autonomous vehicle deployment that tracks regulatory openness and favorable climate at least as much as population density or transport need.
Federal legislative efforts to create a unified national framework had moved slowly through the mid-2020s. The absence of a federal standard created genuine problems for companies trying to scale across state lines, because a safety validation accepted in one state might not satisfy another.
Tesla’s Different Bet and What It Actually Delivered
Tesla occupies a genuinely distinct position in this landscape, and understanding it requires separating the company’s claims from what its technology demonstrably does.
Unlike Waymo or Cruise, Tesla does not use lidar. Musk argued for years that lidar was a crutch and that a camera-based system, trained on data from Tesla’s enormous fleet of consumer vehicles, could achieve full autonomy more cheaply and at greater scale. The approach is defensible in theory: human drivers navigate using eyes, not laser rangefinders, and the sheer volume of Tesla’s fleet data is a genuine competitive asset. Whether it is sufficient remains genuinely contested among researchers.
Tesla’s “Full Self-Driving” (FSD) software, sold as an add-on to consumer vehicles, went through numerous versions over the years. The company updated its architecture over time, with a later version moving to an end-to-end neural network approach that the company argued represented a fundamental improvement. Regulators and safety researchers scrutinized FSD closely, particularly after a series of crashes involving Autopilot, the less capable driver-assistance system that precedes FSD in Tesla’s product hierarchy. The National Highway Traffic Safety Administration opened multiple investigations into Tesla’s automated driving systems.
Tesla unveiled a dedicated robotaxi vehicle, the Cybercab, and has since begun offering robotaxi rides in Texas and Florida, first in modified Model Y SUVs and, from September 2026, in Cybercabs with no steering wheel or pedals in a limited area of Austin. The Cybercab fleet is small, and NHTSA has said it is evaluating the launch, so the scale and regulatory pathway remain open questions. Tesla’s approach, if it works, would represent a fundamentally different economics than sensor-heavy competitors: cheaper vehicles, no lidar, relying on fleet-scale learning. The risk is that cameras alone may not provide the redundancy and precision that other companies argue is necessary for unsupervised operation in complex environments.
The Economics of Running a Driverless Fleet
Even setting aside safety and regulation, robotaxis face an economic structure that is more complicated than early projections acknowledged.
The bull case was simple: remove the driver, who accounts for a large share of ride-hail operating cost, and the unit economics of a taxi trip improve dramatically. Research firms produced projections in the late 2010s suggesting autonomous vehicles could bring ride costs down by half or more compared to human-driven services.
The actual cost picture has proven messier. Driverless vehicles require remote monitoring and intervention capability, which is a labor cost that does not disappear, it just moves off the vehicle. Vehicles need cleaning, maintenance, and repositioning, tasks that still require human workers. The sensor and compute hardware adds vehicle cost. Geofenced operations limit revenue-generating utilization, because a robotaxi that cannot go everywhere cannot serve every trip.
Waymo has not published detailed unit economics, and the company has operated with substantial ongoing investment from Alphabet. The economic viability of robotaxi services at scale has remained, through the mid-2020s, a question rather than a demonstrated fact. Optimists argue that costs will fall as hardware prices drop and fleet scale increases. Skeptics note that the same argument has been made for a decade.
Where This Leaves the Passenger
For someone living in a city where robotaxi service operates today, the experience is stranger and more mundane than either the hype or the fear suggested. Passengers report that Waymo’s vehicles drive competently, if conservatively. They stop at yellow lights that a human driver might run. They occasionally pause at unusual intersections longer than feels necessary. Some riders describe them as uncanny; others simply get on with their journey.
The constraints are real. Service areas are bounded. Surge pricing and availability work similarly to conventional ride-hail. If the vehicle encounters a situation it cannot resolve, a remote operator may take over or the car may pull over and wait for assistance, which can be inconvenient in a way that no promotional video ever quite captures.
For the cities hosting these fleets, the benefits and tensions coexist. Proponents argue that robotaxis could improve mobility for people who cannot drive and could eventually reduce traffic fatalities, which remain in the tens of thousands annually in the United States. Critics raise questions about labor displacement, about data collection by vehicles mapping public streets in detail, and about whether public infrastructure is being privatized without adequate democratic deliberation.
The Honest Forecast
The question is not whether self-driving technology works. It demonstrably does, within defined parameters. The question is how long it takes to expand those parameters to the point where the technology is commercially viable everywhere, and not just in sunbelt cities with good weather and cooperative regulators.
The history of autonomous vehicle timelines suggests specific predictions are hazardous. The pattern has been consistent: technical progress is real, but the gap between a working demonstration and a robust, scalable, economically viable system is vast and tends to be underestimated by the people closest to the technology.
What seems more durable than any specific forecast is the structural observation: the companies that have focused on incremental geographic expansion with rigorous safety validation, rather than promising ubiquitous capability by a particular year, have made more actual progress. Waymo’s cautious market-by-market expansion contrasts with the graveyard of AV programs that collapsed after overpromising. That lesson may be the most important thing the robotaxi industry has produced so far, more valuable in the long run than any particular hardware breakthrough.
The driverless car is real. The driverless car everywhere, for everyone, soon, is still a promise waiting to be earned.