Self-Driving Cars and Robotaxis

The Hardest Robotics Problem

In October 2023, a Cruise robotaxi in San Francisco struck a pedestrian who had been hit by another vehicle and thrown into the autonomous car's path. The robotaxi detected a collision, stopped, then attempted to pull over—dragging the injured pedestrian twenty feet. The incident led to Cruise suspending operations nationwide and losing its California permit.¹

The accident wasn't caused by the autonomous system—a human driver hit the pedestrian first. But it revealed something important: robotaxis must handle situations that never appear in training data, make split-second decisions with incomplete information, and do so under intense public scrutiny. Every mistake is global news.

Self-driving cars are often called the hardest robotics problem in the world. Not because the physics is difficult—wheels on pavement is simpler than bipedal walking—but because the environment is endlessly complex, the stakes are life and death, and the benchmark is comparison to human drivers who have millions of years of evolved neural wetware optimized for navigating the physical world.

And yet: in Phoenix, Arizona, Waymo robotaxis provide over 100,000 paid rides per week. No human is at the wheel. The vehicles operate at night, in rain, and in traffic. They're not perfect—but they appear to be safer than human drivers.²

The world is in an awkward transition period. Self-driving technology works in some places, some of the time, for some use cases. The question is how quickly it will expand—and whether the next few years will see the tipping point that makes autonomy normal rather than novel.

2026 Snapshot — Where Things Actually Stand

The State of the Technology

Autonomy is real but constrained. Commercial robotaxi services operate in:

Phoenix, Arizona: Waymo's most mature market, covering roughly 180 square miles with no safety driver
San Francisco and Los Angeles: Waymo operates commercially in both; LA service opened to the public in 2024
Austin and Atlanta: Waymo service launched in 2025, dispatched through the Uber app rather than Waymo's own
Las Vegas: Zoox, Amazon's purpose-built robotaxi subsidiary, opened public service in 2025
China: Baidu Apollo operates in Wuhan, Beijing, Shanghai, and other cities; Pony.ai and WeRide, both of which listed on Nasdaq in late 2024, hold permits across multiple metros⁷

The field consolidated sharply in this period. General Motors shut down Cruise's robotaxi program in December 2024, folding the remaining engineers into its personal-autonomy and driver-assist work after having spent more than $10 billion on the effort. Cruise had been the clearest second entrant; its exit left Waymo without a scaled American competitor and made the capital requirements of this business unmistakable.

Total robotaxi rides number in the millions annually—significant as proof of concept but tiny compared to ride-hailing (Uber and Lyft complete roughly 10 billion rides per year globally).³

The technology stack has converged on similar approaches:

Multiple sensor modalities (lidar, radar, cameras)
High-definition maps of operating areas
Machine learning for perception and prediction
Redundant compute and vehicle systems
Remote human oversight for edge cases

Tesla's different approach: Camera-only computer vision, trained on data from millions of customer vehicles, without HD maps or geofencing. "Full Self-Driving" is commercially available as a driver-assist feature but requires human supervision. Tesla plans a dedicated robotaxi vehicle for production.

The Safety Question

The data is encouraging but incomplete. Waymo published data claiming their vehicles are significantly safer than human drivers—80% fewer injury-causing crashes, 73% fewer police-reported crashes per million miles.⁴ But:

Sample sizes are still relatively small
Operating domains are carefully selected
Comparison baselines are debated
Rare catastrophic events may not yet have occurred

The fundamental challenge: Proving safety requires millions of miles in diverse conditions. But deploying vehicles to accumulate those miles raises safety concerns. This chicken-and-egg problem has slowed deployment.

Human benchmark: US drivers have roughly one fatal accident per 100 million miles—extraordinarily rare when measured per trip, but totaling 38,000 deaths per year due to enormous aggregate miles driven.⁵ Robotaxis need to be demonstrably better, not just equivalent.

Business Model Uncertainty

The economics remain challenging:

Vehicle cost: $150,000-$200,000+ for a robotaxi (sensors, compute, redundancy)
Utilization: Current fleets are small, limiting economies of scale
Operating costs: Remote monitoring, cleaning, maintenance, charging
Revenue: Comparable to ride-hailing fares in most markets

Waymo (Alphabet subsidiary) has invested well over $5 billion with no public timeline to profitability. Cruise burned roughly $2 billion a year until GM stopped funding it entirely—the clearest evidence available that this business can consume unlimited capital without reaching scale. Tesla's approach, using customer vehicles for training and a cheaper sensor suite, could prove more economical if the camera-only bet pays off.

The unit economics question: Can a robotaxi replace a human driver profitably? Human Uber drivers earn roughly $20-25/hour before expenses. Can autonomous systems deliver rides at lower total cost per mile? The math says yes eventually, but current costs are higher.

Notable Players

Waymo (Alphabet)

Origin: Began as Google's self-driving car project in 2009; became Waymo in 2016.

Approach: Full stack—custom sensors, vehicles (Jaguar I-PACE, Zeekr), software, and commercial operations. Conservative expansion, prioritizing safety data.

Status: The clearest commercial robotaxi success to date, and now effectively unopposed in the United States. Weekly paid rides crossed 100,000 in 2024 and passed 250,000 during 2025, with service in Phoenix, San Francisco, Los Angeles, Austin, and Atlanta, freeway operation added, and further metros announced.⁶

Funding: Backed by Alphabet's resources plus $5.5+ billion in external funding.

Competitive advantage: 15+ years of development, billions of real-world miles, operational experience, and a reputation for safety focus.

Cruise (General Motors) — Wound Down

Origin: Startup founded 2013, acquired by GM in 2016.

Approach: Integrated with GM's manufacturing and vehicle platforms. Operated commercial service in San Francisco until the October 2023 incident.

Status: Defunct as a robotaxi business. Operations were suspended nationwide after the 2023 incident and the California permit was revoked; leadership departed; a limited restart was attempted in 2024. In December 2024, GM ended funding entirely, absorbing the remaining team into its own driver-assistance and personal-autonomy programs.

Why it matters: Cruise is the most instructive failure in the sector, and for a reason that has little to do with the accident itself. A single high-profile incident halted deployment across every market at once—but what killed the company was that the incident arrived while it was still burning $2 billion a year with no path to scale. Robotaxis demand simultaneous excellence in engineering, regulatory relations, and capital discipline. Cruise lost the third and the first two stopped mattering.

Tesla

Origin: Tesla's Autopilot team has developed driver-assist features since 2014; "Full Self-Driving" (FSD) beta began 2020.

Approach: Camera-only (no lidar), no HD maps, trained on billions of miles of data from customer vehicles, aiming for generalization rather than geofencing.

Status: FSD remains a paid option or subscription that still requires driver supervision. Tesla unveiled a dedicated two-seat robotaxi—the Cybercab—in late 2024 and opened a supervised robotaxi pilot in Austin in mid-2025, initially with a company employee in the vehicle. That is a meaningful step, but a supervised pilot in one city is not the general, unmapped autonomy the approach was premised on.

Advantage: Scale of data from millions of customer vehicles; lower hardware cost; vertical integration.

Controversy: Critics argue the camera-only approach is fundamentally limited; regulators have investigated crashes; the timeline for true autonomy has slipped repeatedly, and has done so publicly enough that the slippage is itself now a data point about the approach.

Zoox (Amazon)

Origin: Founded 2014; acquired by Amazon in 2020 for roughly $1.2 billion.

Approach: A purpose-built vehicle with no steering wheel and bidirectional travel, rather than a sensor package retrofitted onto an existing car. This forgoes the option of human takeover entirely—a design commitment, not a feature gap.

Status: Opened public robotaxi service in Las Vegas in 2025, with San Francisco and other markets in testing.

Position: The only remaining well-capitalized American challenger to Waymo, and notable for taking the opposite hardware bet from Tesla.

Chinese Players

Baidu Apollo: Largest Chinese robotaxi operator with commercial service in Wuhan, Beijing, Shanghai, and other cities. Backed by Baidu's AI capabilities and government support.

Pony.ai: Operates in China and previously in the US; partnership with Toyota for robotaxi fleet.

WeRide: Commercial operations in multiple Chinese cities; international expansion.

Didi: Ride-hailing giant developing autonomous technology.

AutoX: Fully driverless operations in Shenzhen.

Advantage: Supportive government policy, rapid infrastructure adaptation, large domestic market.

Trucking-Focused Players

Aurora: Began driverless commercial freight runs on Texas interstates in 2025—the first sustained no-human-in-the-cab trucking operation in the US. Partners with major carriers; passenger applications remain a stated long-term goal.

Kodiak Robotics: Focused on long-haul trucking with human-supervised deployment; also pursuing off-highway and defense applications.

TuSimple: Once the sector's leader, it wound down US operations amid governance and national-security disputes, delisted from Nasdaq, and by the end of 2024 had renamed itself and pivoted out of trucking entirely. A cautionary case in a chapter with several.

Gatik: Middle-mile autonomous delivery for retail (Walmart partnership).

These companies bet that highway trucking (simpler environment, higher economic value) is an easier entry point than urban robotaxis.

The Technical Challenges

The Long Tail Problem

Driving involves countless rare situations: a mattress falling off a truck, a child chasing a ball, a police officer directing traffic with hand gestures, construction zones with improvised signage. Each of these "edge cases" appears rarely but must be handled safely.

The math is daunting: If there are 10,000 distinct edge cases, each occurring once per 100,000 miles, a vehicle must drive a billion miles to encounter each situation just ten times—barely enough data to train reliable responses.

Solutions in progress:

Simulation to generate synthetic edge cases
Generative AI to handle novel situations through reasoning
Conservative fallback behaviors (stop and wait for remote help)
Shared learning across fleets

Perception in Adversity

Sensors work well in good conditions. Challenges multiply in:

Rain, snow, fog: Degrade lidar and camera performance
Direct sunlight: Can blind cameras
Night: Reduced visual information
Dirty sensors: Road grime accumulates
Sensor failures: Redundancy is essential but adds cost

Each condition requires specific engineering and testing. The number of combinations (heavy rain at night in a construction zone) multiplies complexity.

Prediction and Interaction

Other road users are unpredictable. A robotaxi must predict:

Will that pedestrian step into the street?
Is that car going to run the red light?
What is that cyclist signaling?
Is that stopped vehicle going to door?

Humans use subtle cues—body language, eye contact, vehicle behavior patterns—that are difficult to encode in software. Getting this wrong leads to either dangerous situations (assuming the pedestrian will stop) or overly cautious behavior (stopping unnecessarily, frustrating passengers).

The Mapping Problem

Most autonomous systems rely on high-definition maps—detailed 3D representations of road geometry, lane markings, signals, signs, and landmarks. These maps must be:

Created initially for any new operating area
Updated continuously as roads change
Accurate to centimeters

This requirement limits scalability: expanding to a new city requires months of mapping before operations can begin.

Tesla's approach—no HD maps, pure computer vision—could solve this but requires the AI to generalize from learned patterns rather than matching to known maps. Whether this achieves comparable safety remains debated.

Validation: How Safe Is Safe Enough?

The fundamental challenge: How do you prove a vehicle is safe before deploying it?

Simulation can test billions of scenarios but may miss edge cases that don't appear in the simulation.

Controlled testing is necessary but can't replicate all real-world conditions.

Real-world deployment generates the best data but risks actual harm.

Statistical validation requires enormous sample sizes. To prove a vehicle is safer than human drivers with 95% confidence might require billions of miles without serious incidents.

Current approaches combine all methods: extensive simulation, controlled testing, limited deployment with monitoring, and gradual expansion as safety data accumulates.

The Path to Mass Deployment

Near-Term Likely (2026-2032)

Robotaxi expansion in favorable markets: Waymo and competitors will expand to additional US cities with good weather, clear road markings, and supportive regulation. Coverage will remain geofenced to mapped areas.

Hundreds of thousands to low millions of robotaxis globally by the early 2030s, concentrated in the US, China, and select other markets.

Autonomous trucking commercializes: Highway trucking, with its more predictable environment and strong economic incentive, will see commercial deployment. Initially on specific routes (highway-to-highway) with human drivers handling first and last miles.

Driver-assist becomes more capable: Features like Tesla's FSD, GM's Super Cruise, and competitors will handle more driving situations with less supervision, blurring the line toward autonomy.

Regulatory frameworks stabilize: Federal standards for autonomous vehicle testing and deployment will emerge in the US; China will continue rapid regulatory adaptation; Europe will develop unified standards.

Key milestone: Robotaxi rides numbering in hundreds of millions annually; clear statistical evidence of safety improvement over human drivers.

Plausible (2032-2040)

Geographic expansion accelerates: Improved AI (potentially leveraging large language models for reasoning) enables operation in more challenging environments—rain, snow, complex urban areas.

Mapping constraints ease: Better generalization reduces dependence on HD maps; new areas can be opened faster.

Cost reductions: Sensor costs decline; vehicle designs optimize for autonomy; utilization increases; robotaxi rides become cheaper than traditional taxis.

Mixed autonomy fleets: Some vehicles fully autonomous, others with remote supervision or occasional human backup. The distinction blurs.

Personal vehicles gain autonomy: Consumers can purchase vehicles that are genuinely self-driving in many conditions, not just driver-assist.

Ownership model shifts: Vehicle-as-a-service becomes viable in dense areas; some households reduce from two cars to one-plus-robotaxi.

Key milestone: Robotaxis available in most major cities worldwide; total cost per mile competitive with or cheaper than private vehicle ownership.

Wild Trajectory (2040+)

Human driving becomes rare: For most trips, in most conditions, autonomous vehicles are so much safer, cheaper, and more convenient that human driving is a niche activity—like riding horses.

Vehicle design transforms: Without human drivers, vehicles can be redesigned. Mobile offices, sleeping pods, entertainment centers. "Cars" become rooms that move.

Parking becomes obsolete: Vehicles drop off passengers and circulate or park remotely. Urban real estate reconfigures.

The second-order cascade:

Auto insurance shrinks dramatically (fewer accidents, clearer liability)
Emergency rooms see fewer trauma cases
Alcohol becomes irrelevant to transportation
The elderly and disabled gain mobility independence
Traffic enforcement largely disappears
Urban planning is unconstrained by parking requirements

Second-Order Effects

Urban Form

Parking transformation: In US cities, parking often consumes 30% of land area. If vehicles can drop passengers and park remotely or circulate, this land becomes available for housing, parks, or commercial use.

Traffic flow improvement: Autonomous vehicles can coordinate—accelerating and braking in platoons, optimizing intersection timing, eliminating the "phantom traffic jams" caused by human overreaction.

Congestion uncertainty: Will robotaxis reduce congestion (better utilization, coordinated flow) or increase it (more vehicle-miles as cheap rides replace transit and walking)?

Suburban accessibility: Self-driving may make distant suburbs more attractive (commute time becomes productive time) or make dense urbanism more viable (reduced need for parking).

Labor Markets

Driving jobs at risk: Roughly 3 million professional truck drivers in the US; hundreds of thousands of taxi and ride-share drivers; delivery drivers, bus drivers, and others. Not all jobs disappear—local delivery, specialized transport, and oversight roles persist—but displacement will be significant.

Transition timeline: Mass displacement likely takes 10-15 years rather than happening overnight. New jobs emerge (fleet management, vehicle maintenance, remote monitoring). The question is whether transition is managed or disruptive.

Geographic concentration: Truck driving is one of the most common jobs in many US states. Concentrated job loss could devastate certain regions.

Safety and Public Health

Lives saved: If autonomous vehicles are 80% safer than human drivers, widespread adoption could prevent 30,000+ US deaths annually—more than from homicide.

Emergency services: Fewer accidents means fewer emergency room visits, reduced demand for emergency responders, and changes to healthcare systems.

Impaired driving eliminated: One of the leading causes of traffic deaths becomes impossible.

Accessibility

Mobility for all: The elderly, disabled, visually impaired, and those unable to drive gain independent mobility. This could transform quality of life for tens of millions.

The child transport problem: Parents spend enormous time transporting children. Self-driving vehicles could eventually enable independent child mobility (with appropriate safeguards).

Economic Transformation

Cost of mobility: If robotaxi rides cost $0.50-1.00 per mile (plausible with high utilization and no driver costs), transportation becomes dramatically cheaper for non-car-owners.

Vehicle utilization: Private cars sit idle 95% of the time. Robotaxis might operate 50-70% of the time. The same transportation could be delivered with far fewer vehicles.

The end of car ownership: Some predict most urban households won't own cars—paying for mobility as needed rather than maintaining depreciating assets. This would transform automotive manufacturing, insurance, parking, and maintenance industries.

Risks and Guardrails

Safety Risk

Insufficiently tested systems: Pressure to deploy before technology is ready could lead to accidents that harm both passengers and the industry's reputation.

The 1,000x problem: A system that's 99% as safe as human drivers still means thousands of preventable deaths. The public may not accept "almost as safe."

Guardrails: Conservative expansion policies, extensive simulation before deployment, real-time monitoring, immediate halt when problems emerge, regulatory oversight with meaningful enforcement.

Cybersecurity

Vehicle hacking: An autonomous vehicle is a computer on wheels. Compromised vehicles could cause accidents, be used for crimes, or be disabled en masse.

Fleet attacks: Unlike individual vehicles, robotaxi fleets could be targeted simultaneously—potentially halting transportation in a city.

Guardrails: Security-first design, redundant systems, ability to take vehicles offline, ongoing security monitoring and updates.

Liability and Insurance

Who is responsible for accidents? The vehicle manufacturer, software developer, fleet operator, remote monitor, or some combination? Current liability frameworks don't clearly answer this.

Insurance transformation: If accidents decrease, premiums should decrease—but until the statistics are clear, insurers may price conservatively.

Guardrails: Clear regulatory frameworks for liability; mandatory insurance requirements; data recording requirements for accident reconstruction.

Equity and Access

Geographic disparities: Robotaxis may deploy in wealthy cities first, bypassing rural areas and smaller cities for years or decades.

Cost access: Will robotaxis be affordable for all, or a premium service? Early pricing in some markets exceeded Uber fares.

Digital divide: Ordering a robotaxi requires a smartphone and payment method—barriers for some populations.

Guardrails: Requirements for service in underserved areas; subsidized access programs; integration with public transit.

Privacy and Surveillance

Constant recording: Autonomous vehicles have cameras and sensors continuously capturing their environment. This data could be used for surveillance.

Trip data: A complete record of everywhere you go, when, and with whom. Law enforcement, advertisers, and others may want access.

Guardrails: Data minimization policies; restrictions on data sharing; clear user consent; regulatory limits on surveillance use.

Transition Disruption

Job displacement: Millions of driving jobs at risk over the next 10-20 years. Without managed transition, this could cause significant economic hardship.

Community impact: Truck stops, motels, and towns along highways depend on driver traffic. Gas stations, repair shops, and related businesses face disruption.

Guardrails: Extended transition timelines; retraining programs; support for affected communities; gradual phase-in rather than sudden displacement.

The AI Acceleration Factor

Self-driving represents a domain where AI acceleration could be dramatic:

Foundation models for driving: Large language models demonstrate broad reasoning capabilities. Applying similar approaches to driving—training on vast amounts of driving data and reasoning about novel situations—could address the long-tail problem.

Simulation at scale: AI can generate synthetic driving scenarios, test systems against millions of edge cases, and identify failures before deployment.

Rapid iteration: Traditional automotive development takes 5-7 years per vehicle generation. Software-defined vehicles can improve continuously through over-the-air updates.

Cross-fleet learning: Every mile driven by any vehicle in a fleet can train all other vehicles. A single unusual encounter teaches the entire fleet.

The wildcard: Current progress has been slower than optimistic predictions (2015-era forecasts expected widespread autonomy by now). But AI capabilities have advanced dramatically. The gap between current systems and human-level driving may close faster than linear extrapolation suggests.

The Path Forward

Self-driving cars have been "almost here" for a decade. The technology has improved enormously but full deployment has remained elusive. This pattern—perpetual imminence—has bred skepticism.

But the skeptics may be wrong about the timeline even if they were right about the pace. Capabilities that seemed far off in 2015 are now operational in limited domains. The gap between current performance and human-level driving is narrower than ever.

The next few years will be decisive:

If safety data continues to improve, and robotaxis demonstrate statistically significant safety advantages over human drivers, regulatory barriers will fall and deployment will accelerate.

If AI reasoning capabilities transfer to driving, the long-tail problem may become solvable, enabling operation in more challenging environments without years of domain-specific training.

If costs decline through sensor commoditization, vehicle optimization, and high utilization, the economic case becomes irresistible.

The end state—a world where human driving is rare, accidents are infrequent, and mobility is a service rather than an asset—may arrive gradually and then suddenly. The transition will be neither as fast as boosters predict nor as slow as skeptics assume.

What's clear is that this is not a question of if but when—and whether society manages the transition wisely.

Endnotes — Chapter 14

The October 2023 Cruise incident in San Francisco was extensively documented in regulatory filings, media reports, and subsequent investigations. DMV revoked Cruise's permit; operations were suspended nationwide.
Waymo safety data from company reports and filings. The company publishes regular safety reports comparing their crash rates to human driver baselines in their operating domains.
Uber reported approximately 2.6 billion trips in Q4 2023 alone. Combined with Lyft, DiDi, and other platforms, global ride-hailing trips exceed 10 billion annually.
Waymo published comparative safety data in 2024 showing 80%+ reduction in injury-causing crashes compared to human drivers in their operating area. Independent verification is ongoing.
NHTSA data shows approximately 1.3 fatalities per 100 million vehicle miles traveled in the US. With roughly 3.2 trillion miles driven annually, this results in approximately 38,000-42,000 deaths per year.
Waymo ride data from company announcements. The 100,000+ weekly rides in Phoenix as of late 2024 represents the largest robotaxi deployment in the US.
Chinese robotaxi deployments documented by company announcements and regulatory approvals. Wuhan has become one of the largest robotaxi markets globally.