Autonomous Vehicles: what are the biggest unsolved problems?
In our autonomous vehicle market deck, you will find everything you need to understand the market
SUMMARY
Autonomous Vehicles: what are the biggest unsolved problems? The biggest one is generalization: autonomous driving already works impressively well inside selected operating areas, but nobody has shown that the same performance can travel cheaply and reliably across most roads, climates, countries and unusual situations.
The old question of whether a car can drive through a complicated city with nobody behind the wheel is increasingly settled. Waymo's driverless mileage, commercial ride volumes, Pony.ai's growing fleet and Zoox's public service show that Level 4 autonomy is already a real transportation product in a small number of places.
The harder problem begins when familiar things appear in unfamiliar combinations. Heavy rain, temporary markings, emergency vehicles, cyclists, construction workers and sensor degradation are each manageable on their own; combining several of them produces the long tail that remains difficult to anticipate and test.
More road mileage alone cannot close that gap. Once a dangerous failure becomes extremely rare, even hundreds of millions of additional miles can provide surprisingly little statistical certainty about whether a software update has actually reduced the risk.
That pushes autonomous-driving validation toward a mix of real roads, closed tracks, simulation and deliberately generated scenarios. The awkward part is that engineers are very good at reproducing failures they already understand and much less able to test the important failure nobody has imagined yet.
Weather, maps and human behavior all point back to the same generalization problem. A scalable autonomous driver has to keep working when visibility deteriorates, stored road information becomes stale, or another road user behaves ambiguously rather than following a clean textbook trajectory.
Safety evidence is becoming unusually strong inside deployed operating areas. Waymo's large driverless dataset reports substantially fewer injury and serious crashes than comparable human driving, but results from Phoenix, San Francisco or Atlanta still cannot prove equivalent performance on snowy mountain roads or in countries with very different driving habits.
Remote assistance is another useful test of whether autonomy is genuinely scaling. Human support is not secretly driving the cars, but the ratio of support workers to vehicles matters: if fleets can grow much faster than remote staffing, the operating model becomes far more convincing.
Economics may ultimately become as important as the remaining technical problems. Positive unit economics in selected robotaxi zones are encouraging, yet hardware, empty repositioning, charging, cleaning, maintenance, insurance, remote operations and enormous R&D spending still separate a successful service from a profitable global transportation network.
The clearest benchmark for progress is therefore how much preparation the next city requires. When an autonomous driver can enter a very different place with dramatically less mapping, testing and local engineering than the previous one, handle strange situations it has never seen and do it at competitive cost, general self-driving will be much closer to solved.
This market map, featured in our autonomous vehicle market deck, highlights top companies and startups in the autonomous vehicle market
Has autonomous driving actually been solved anywhere?
Yes — autonomous driving currently works well enough to run large driverless transportation services in a handful of cities.
Waymo offers the clearest proof. Its latest safety analysis covered more than 220 million fully autonomous miles across five U.S. operating areas. Compared with human drivers on similar roads, the Waymo Driver recorded 82% fewer crashes involving an injury, 82% fewer crashes with an airbag deployment and 94% fewer crashes causing serious or fatal injuries. The company is also driving more than four million autonomous miles a week.
The commercial side has moved well beyond tiny pilots too. Waymo was already providing more than 400,000 rides a week earlier this year. Pony.ai's latest quarterly results show its robotaxi fleet reaching 1,975 vehicles, while robotaxi-service revenue jumped from $1.5 million to $12.1 million year over year. Zoox has opened its purpose-built robotaxi to public riders in Las Vegas.
So one old question can mostly be retired: can a car drive around a complicated city with nobody behind the wheel? Yes, it can.
The remaining problem is much bigger. Nobody has shown that an autonomous driver can take the performance achieved in selected cities and reproduce it across most roads, weather conditions, countries and unusual situations at a cost that makes mass deployment worthwhile.
| What autonomous vehicles can do today | What remains unproven |
|---|---|
| Run fully driverless urban trips | Drive almost anywhere |
| Operate commercial robotaxi fleets | Scale cheaply across thousands of cities |
| Beat human crash benchmarks in selected operating areas | Beat humans across very different driving environments |
| Handle millions of routine trips | Reliably cover the rarest combinations of events |
If you want more recent data on this point, please see our latest autonomous vehicle market report.
Are rare edge cases still the biggest autonomous-driving problem?
Yes — rare edge cases are still the hardest technical problem in autonomous driving today because a road can produce combinations that nobody has explicitly trained or tested.
Routine driving is increasingly manageable. Cars follow lanes, maintain distance, react to traffic lights and navigate ordinary intersections millions of times. The difficulty starts when familiar elements combine in unfamiliar ways.
Imagine temporary lane markings during heavy rain, a police officer overruling the traffic light, an ambulance blocking half the intersection and a cyclist suddenly entering the remaining lane. Every element is understandable on its own. The vehicle still has to interpret the combination correctly within seconds.
Ten unusual weather conditions combined with ten strange road layouts, ten kinds of temporary obstruction, ten abnormal road-user behaviors and ten sensor problems already create 100,000 combinations. Add different speeds, lighting, distances and timings and the number of possible situations explodes.
Machine learning has made autonomous vehicles much better at generalizing beyond exact training examples. We still do not know how far that generalization can stretch before performance deteriorates.
As this chart shows, and as featured in our autonomous vehicle market deck, search interest in autonomous vehicles has continued to rise
Why can't self-driving cars solve edge cases by collecting more miles?
More driving data helps, but autonomous vehicles cannot simply drive their way through every rare situation.
Suppose a dangerous failure appears once every 100 million miles. If engineers release new software and want to know whether that failure rate has been cut in half, another few million miles tells them almost nothing. Even hundreds of millions of miles may produce too few examples for a clean statistical answer.
This is why autonomous-driving companies increasingly combine public-road mileage with simulation, closed-course tests and deliberately generated scenarios. ISO 34505, published in 2025, formalized methods for evaluating autonomous-driving test scenarios using factors such as frequency, criticality and complexity. Other standards under development are going further into virtual-test environments and scenario description.
The difficult part is deciding whether those tests represent reality well enough. Engineers can recreate a crash they already know about and run thousands of variations. They cannot easily simulate an important failure nobody has imagined yet.
Can self-driving cars handle heavy rain, snow and fog today?
Self-driving cars can already operate through some difficult weather, but genuinely all-weather autonomous driving is still unsolved.
Pony.ai says its current commercial fleet has continued operating through heavy rainstorms in major Chinese cities, which shows how far the systems have come. Waymo has also been deliberately testing in colder and snowier environments as it expands beyond the climates where robotaxis first reached scale.
The difficult cases start when several things deteriorate together. Cameras lose contrast in fog, glare or heavy rain and can become obscured by water or snow. LiDAR can suffer from scattering and weaker returns. Radar generally handles poor visibility better but gives the system less visual detail.
Then the road itself changes. Snow can hide curbs and lane markings. Standing water produces reflections. At night, wet pavement turns headlights into visual noise. Accumulated snow may erase landmarks that normally help localization.
A vehicle does not need perfect visibility to drive safely; humans certainly do not have it. But an autonomous system has to recognize how much information it has lost, slow down appropriately and know when conditions have crossed its operating limit.
| Technology | What it does well | What bad weather can disrupt |
|---|---|---|
| Cameras | Reads visual detail, signs and road users | Fog, glare, darkness, rain, blocked lenses |
| LiDAR | Builds precise 3D geometry | Rain, fog and airborne snow |
| Radar | Measures distance and speed through tougher weather | Provides less detailed scene information |
| Maps and localization | Gives precise prior knowledge of the road | Road changes, snow coverage and positioning errors |
| Sensor fusion | Lets sensors compensate for each other | Several sensors can degrade at the same time |
This chart, included in our autonomous vehicle market deck, illustrates yearly VC funding for autonomous vehicle startups
Can autonomous vehicles tell what pedestrians and cyclists are about to do?
Autonomous vehicles can detect people very well these days, but predicting what those people will do next remains much harder.
A pedestrian standing beside a crossing might step into the road, wave a vehicle through or stay on the pavement. A cyclist looking over one shoulder may be about to merge. A car edging out of an alley can be yielding slowly or trying to claim the gap.
Human drivers make these judgments constantly from tiny movements, speed changes, eye contact and local habits. Much of road behavior is an informal negotiation.
Prediction models can assign probabilities to several future trajectories, but the other person is also reacting to the autonomous vehicle. If the robotaxi inches forward, the pedestrian may stop. If it waits, another driver may interpret that as permission to go.
Dense traffic makes this especially awkward. A vehicle that refuses to move until every ambiguity disappears can become excessively cautious. One that tries too hard to imitate assertive human driving creates the opposite problem.
Why do construction zones and emergency scenes still confuse autonomous vehicles?
Construction zones and emergency scenes remain difficult for autonomous vehicles because both can temporarily overturn the normal road rules.
A mapped lane may suddenly disappear. Old markings remain visible underneath temporary paint. Cones create a route that cuts across the usual traffic pattern. A worker or police officer can wave cars through a red light or across a double yellow line.
That leaves the autonomous vehicle choosing between competing instructions. Should it trust its map, the painted lane, temporary cones, a portable sign or a person's hand gesture? Humans often understand immediately that somebody with authority has temporarily changed the rules.
Emergency scenes add another layer. Fire engines may block most of a junction, ambulances can approach from unusual directions and officers may tell vehicles to stop somewhere normally prohibited.
A review of 74 studies on autonomous vehicles and first responders published in 2025 found that this interaction had received surprisingly little research attention relative to its importance. California's new autonomous-vehicle rules now require updated first-responder interaction plans, manual override access and two-way communication capable of reaching remote support quickly.
Autonomous-driving developers have become much better at detecting temporary objects and road changes. The remaining difficulty is understanding which normal rules no longer apply and which temporary instruction takes priority.
This chart, included in our autonomous vehicle market deck, shows how Waymo is winning in autonomous vehicles
Do self-driving cars still need detailed maps?
Detailed maps remain useful for self-driving cars today, although the industry's direction is clearly toward reducing dependence on them.
High-definition maps can contain lane geometry, curbs, crossings and landmarks at far greater precision than ordinary navigation maps. The autonomous system gets a prior picture of what the road should look like and can compare its sensors against that picture.
The weak point appears once fleets expand. Mapping every road takes work. Keeping every road current takes even more. Construction, repainted lanes, temporary closures and new intersections constantly make stored information stale.
That is why researchers and autonomous-driving developers have been pushing harder toward lightweight, implicit and map-free approaches. A system that can understand an unfamiliar road directly is much easier to scale than one that needs detailed preparation before entering every new area.
Maps are unlikely to disappear completely because prior knowledge is valuable. The real test is whether the car remains safe when the map is missing or wrong.
How does a self-driving car know when it is confused?
Knowing when it is confused is still one of the most important unsolved problems for a self-driving car.
A wrong prediction made with high confidence is especially dangerous.
An autonomous vehicle therefore needs some estimate of uncertainty around perception, prediction and planning. Sensors may disagree. A road layout may look unlike previous examples. Localization can become unreliable. Another road user may behave in a way that fits several interpretations.
Humans naturally change behavior when confidence drops. We slow down in heavy rain, leave more room around somebody driving strangely and sometimes stop when we cannot understand a junction.
Encoding the same instinct into an autonomous vehicle creates trade-offs. If the car becomes cautious whenever uncertainty rises slightly, passengers get a painfully slow service. If it stops in the wrong place, caution itself can create danger. If it calls remote assistance every few minutes, the operating model becomes expensive. That gets awkward fast.
If you want more recent data on this point, please see our latest autonomous vehicle market report.
This chart, included in our autonomous vehicle market deck, illustrates yearly funding for autonomous vehicle startups
Can we really prove that self-driving cars are safer than humans?
We can already show that some self-driving systems are safer than humans inside specific operating areas, but we cannot turn that into a universal claim about autonomous driving.
Waymo's latest analysis is the strongest large-scale evidence available. Across more than 220 million fully autonomous miles, it reported 82% fewer injury crashes and 94% fewer serious-or-fatal-injury crashes than comparable human driving. Atlanta is particularly useful because it was a newer operating environment: over 5.4 million driverless miles there, Waymo reported 86% fewer injury crashes and no serious-or-fatal-injury crash where the human benchmark would have predicted roughly 1.2.
That gives us much more than a demo or a company saying its software feels safe. We have a large real-world dataset showing a substantial safety advantage in several cities.
The limit appears when the operating environment changes. Evidence from Phoenix, San Francisco or Atlanta does not prove the same performance on snowy mountain roads, narrow European streets or roads in countries with very different driving habits.
Validation also gets statistically harder as crashes become rarer. Proving that an already safe system became another 20% safer may require huge mileage.
This is why current safety work increasingly mixes road data with simulations, targeted scenarios, closed tracks and structured safety cases. California's latest rules go in the same direction: companies moving through testing toward commercial deployment now have to submit structured safety cases rather than relying only on miles driven.
| Evidence | What we learn from it | Main limitation |
|---|---|---|
| Real driverless mileage | How deployed vehicles actually perform | Rare crashes require huge exposure |
| Human-driver comparisons | Whether AV crash rates are better or worse | Comparison depends on road and operating area |
| Simulation | Lets engineers run huge numbers of difficult cases | Simulated reality can miss unknown failures |
| Closed-track tests | Makes dangerous scenarios repeatable | Real roads are much messier |
| Safety cases | Connects different forms of evidence | Still depends on the quality of the underlying evidence |
Will end-to-end AI solve self-driving?
End-to-end AI could make self-driving much more general, but today it creates a difficult trade-off between better learning and easier safety validation.
Older autonomous-driving stacks separate tasks such as perception, prediction and planning into clearer modules. Increasingly, developers are using larger learned models that can connect more of the driving task directly and absorb patterns from enormous datasets.
The attraction is straightforward. Engineers cannot hand-code a rule for every strange situation a vehicle will encounter. A strong learned model may generalize to combinations nobody explicitly programmed.
Validation becomes harder when the behavior of the system is spread across billions of learned parameters. With a modular stack, engineers can sometimes isolate a broken perception rule or planning decision. With a large end-to-end model, explaining exactly why one unusual scene produced one unusual maneuver can be much harder.
Autonomous driving therefore needs better average performance, rare-case reliability and a credible way to show that a new model has not quietly introduced another failure.
This chart, included in our autonomous vehicle market deck, compares the main business model options for autonomous trucking companies
Do remote humans secretly drive robotaxis?
No — remote humans do not secretly drive Waymo robotaxis, but autonomous fleets still use human support when a vehicle wants extra context.
Waymo published unusually specific figures earlier this year. Around 70 remote-assistance agents were on duty worldwide at a given time while the company had roughly 3,000 vehicles operating. Those cars were completing more than four million miles and 400,000 rides each week.
That works out to roughly one available remote-assistance worker for every 43 vehicles. More importantly, Waymo says these workers provide information rather than steering or braking the cars. The autonomous system can accept or reject the advice.
A common example would be a vehicle seeing an ambiguous temporary road closure and asking whether a particular route appears open. The human helps interpret the scene; the driving system still chooses and executes the maneuver.
The number to watch is human support per vehicle. If fleets grow 100-fold while remote staffing barely moves, autonomy scales. If support staff have to grow almost as quickly as the number of cars, the business loses part of the advantage it was supposed to create.
Can robotaxis actually become profitable?
Robotaxis can already show positive economics in selected operations, but nobody has proved that large autonomous fleets can generate strong company-level profits at global scale.
Pony.ai gives us one of the freshest commercial data points. Its latest quarterly robotaxi-service revenue reached $12.1 million, up 691% year over year, while fare-charging revenue rose even faster. The company says its newer robotaxis have achieved positive unit economics in its leading Chinese markets.
That is an important milestone because revenue is increasingly coming from paying rides rather than demonstrations. But Pony.ai still spent $56.2 million on R&D during the same quarter, more than four times its robotaxi-service revenue. Its total operating expenses were $72.1 million.
Waymo shows the capital intensity from another angle. The company raised $16 billion earlier this year after providing 15 million rides during 2025. A business capable of serving millions of customers can therefore still require enormous amounts of capital while expanding.
Fleet economics will depend heavily on utilization. A robotaxi generates no fare while repositioning empty, charging, being cleaned or sitting idle. Expensive sensors, computing hardware, maintenance, insurance and remote operations also replace part of the labor cost removed by eliminating the driver.
McKinsey previously estimated fully burdened costs of about $8.20 per vehicle-mile for a 1,000-car U.S. fleet, with mature large-scale economics potentially approaching $1.30 per mile. The gap between those two numbers is huge. Real fleets still need to travel most of that cost curve.
| Cost problem | Why robotaxi operators care |
|---|---|
| Vehicle and autonomous hardware | Expensive cars need many productive miles |
| Empty driving | Adds mileage without adding revenue |
| Charging and cleaning | Takes cars out of service |
| Maintenance | Complex sensors and compute add servicing costs |
| Remote operations | Keeps some human labor in the system |
| R&D | Can dwarf the revenue of an early commercial fleet |
If you want more recent data on this point, please see our latest autonomous vehicle market report.
This chart, featured in our autonomous vehicle market deck, shows the share of revenue generated by each customer segment in the autonomous vehicle market
Can autonomous vehicles expand to any city?
Autonomous vehicles are entering new cities much faster now, but nobody has demonstrated cheap, repeatable expansion across almost any geography.
Every city changes the problem. Atlanta does not drive like San Francisco. London adds left-side traffic and very different street geometry. Tokyo brings another set of lane conventions and road behaviors. Snowier markets stress perception and localization in ways that Phoenix cannot.
The useful metric here is how much extra engineering each new city requires. If deployment still needs years of local tuning, the technology behaves like a collection of city-specific systems. If city number 30 needs a fraction of the work required for city number three, we are getting much closer to a general driver.
Waymo's recent expansion is encouraging because the company has moved beyond a small cluster of similar markets and is preparing or operating across a much wider mix of U.S. and international cities. Pony.ai is also expanding abroad through partnerships and says its latest deployment agreements cover thousands of robotaxis.
Lower-density areas introduce another problem: even if the driving is easier, the economics can be worse. A robotaxi in a dense city can find another passenger nearby. A rural vehicle may drive a long distance empty between rides. Weak lane markings, animals, unpaved roads and poor connectivity add technical complications too.
Highway trucking may scale differently because highways are more structured and freight routes repeat. Yet terminals, construction zones and local pickup or delivery bring back many of the difficult situations.
Can autonomous vehicles be kept safe from hackers?
Cybersecurity will remain a permanent safety problem for autonomous vehicles because these cars are connected computers that can physically move at road speed.
NHTSA has highlighted threats including GPS spoofing, manipulation of road signs, radar or LiDAR interference, camera blinding and attempts to trigger machine-learning errors. Those sit alongside conventional software risks involving networks, cloud systems, wireless updates and suppliers.
Fleet scale makes the problem more serious. A remotely exploitable software vulnerability could potentially be attacked across many connected vehicles quickly.
Centralized software also gives autonomous fleets an advantage. Operators can watch for anomalies across vehicles and deploy fixes much faster than a traditional mechanical recall.
Autonomous vehicles therefore need strong compartmentalization, authenticated communication, attack detection and the ability to reach a safe state even when part of the system has been compromised.
This chart, included in our autonomous vehicle market deck, shows how robotaxi platform technology has evolved over time
Are regulation and liability still blocking autonomous vehicles?
Regulation and liability slow autonomous-vehicle deployment today, but neither looks like the main reason general self-driving remains unsolved.
California's new AV rules show how the legal framework is maturing. Companies moving toward commercial deployment must progress through testing stages, meet mileage requirements and prepare structured safety cases. The rules also create procedures for autonomous-vehicle traffic violations and require stronger first-responder communication.
NHTSA continues collecting crash reports from operators of automated-driving systems through its Standing General Order. Its public data now runs into 2026, although the agency itself warns that reporting differences, duplicate reports and data-quality issues make raw counts difficult to compare across companies.
Liability is becoming clearer as vehicles remove the human driver. When an autonomous-driving system is responsible for the driving task, attention naturally shifts toward the operator, manufacturer, software, sensors and fleet-maintenance process.
The messy cases sit between those categories. A crash might begin with a dirty sensor, an incorrect map, a software decision or unusual behavior from another driver. Courts, insurers and regulators still have to decide how responsibility should be divided.
If you want more recent data on this point, please see our latest autonomous vehicle market report.
Why are people still scared of self-driving cars?
People remain surprisingly nervous about self-driving cars even as real-world safety evidence gets stronger.
AAA's latest nationwide survey found that only 13% of U.S. drivers would trust riding in a self-driving vehicle, while roughly six in ten said they were afraid. The improvement from the previous year was small. AAA also found that 53% would avoid riding in a robotaxi even though nearly three quarters knew what robotaxis were.
The contrast with operating data is becoming hard to ignore. As seen above, Waymo's large driverless dataset shows substantially fewer injury crashes than comparable human driving, yet public trust remains low.
People clearly do not judge autonomous and human mistakes the same way. A human crash is familiar. A driverless car behaving strangely at an intersection is unusual enough to circulate online and can shape opinions far beyond the actual frequency of the event.
The Cruise episode demonstrated how quickly confidence can collapse. After a pedestrian was dragged by a Cruise vehicle following an initial collision with another car, California suspended the company's driverless permits. The regulatory fallout and subsequent retreat of Cruise's robotaxi operation showed that one badly handled safety episode can damage far more than one deployment.
For autonomous vehicles, matching human safety may therefore be too weak a commercial target. The systems probably need to become visibly better than humans while avoiding the bizarre failures people remember.
In our autonomous vehicle market deck, we identify pain points entrepreneurs should prioritize
Will ordinary people be able to buy a fully self-driving car soon?
A personally owned car that can drive almost anywhere without supervision remains much harder than today's robotaxis, and nothing currently deployed has solved that problem.
A robotaxi operator chooses where its cars work. It can block roads or weather conditions the system does not support. Vehicles return to managed depots, sensors can be inspected regularly and fleet staff are available when something unusual happens.
Private cars live a much messier life. The owner may drive from a warm city into mountain snow, take an unfamiliar rural road at night, ignore a dirty sensor for weeks or lose cellular coverage far from a service area.
That is a radically broader requirement than running a Level 4 robotaxi inside a defined operating domain.
Current consumer systems avoid that burden by keeping a human responsible or restricting when automated driving can be used. A genuinely driverless private vehicle would need to cope safely with a much larger slice of the road world without a fleet operator choosing the easy boundaries.
So what are the biggest unsolved problems in autonomous vehicles today?
The biggest unsolved autonomous-vehicle problem today is generalization: making the extraordinary performance already achieved in selected places survive when the road, weather, country and situation change.
The long tail comes first. Autonomous systems still need to handle rare combinations of temporary infrastructure, unusual human behavior, emergency instructions, bad weather and sensor degradation without engineers anticipating each combination beforehand.
Safety validation comes immediately after it. Once serious failures become extremely rare, proving that a software update is safer than the previous one becomes statistically difficult. Simulation and scenario-based testing can fill part of that gap, but the industry still has to show that those tests capture the failures nobody already knows to look for.
Then comes economics. Driverless fleets need to turn cheaper hardware, higher utilization and lower human-support requirements into durable profits. The latest robotaxi revenue growth shows that commercialization is finally moving fast, but the gap between positive unit economics in selected zones and a profitable global transportation network remains enormous.
Weather, maps, human-intent prediction, cybersecurity, regulation and public trust all sit inside that larger challenge.
Autonomous driving has moved further than the old debate suggests. We already have cars carrying paying passengers without drivers, large safety datasets and fleets expanding into increasingly different cities. The remaining leap is making that performance portable.
Until an autonomous driver can enter a new place with dramatically less preparation, handle the weird situations it has never seen and do so cheaply enough to beat human-driven transport, general self-driving remains unsolved.
If you want more recent data on this point, please see our latest autonomous vehicle market report.
This chart, included in our autonomous vehicle market deck, shows the share of revenue by region across Europe, Asia, North America, Africa, and South America in the autonomous vehicle market
OUR METHODOLOGY
There is no single accepted test for deciding whether autonomous driving has been solved. We therefore broke the question into the main dimensions that determine whether autonomous driving can move from successful Level 4 deployments in selected areas to a genuinely general transportation technology.
For each dimension, we looked for recent evidence that could tell us something concrete about where autonomous vehicles stand today. We prioritized real driverless mileage, crash outcomes, commercial fleet activity, operating economics, deployment experience, regulatory requirements and current technical standards. Research and technical publications were used where real-world operating data alone could not answer the question.
We kept different kinds of evidence separate. A large robotaxi fleet can demonstrate that autonomous driving works inside a defined operating environment, while it says much less about performance on unfamiliar roads, in different weather or in another country. Positive unit economics in one operating area also answers a different question from company-wide profitability or the cost of expanding into dozens of new markets.
Where comparisons were possible, we favored like-for-like evidence. Safety results received more weight when autonomous vehicles were compared with human drivers exposed to similar roads and operating conditions. Company disclosures were used for their own fleet, revenue and operational figures, while regulators, standards bodies and independent research provided broader reference points.
The final ranking comes from aggregating those signals across the different dimensions and asking which unresolved issues still constrain wider deployment most strongly. Problems that affect many environments, make safety harder to prove or materially change the economics of expansion carried more weight than difficulties already handled reasonably well inside established operating areas.
Key sources include Waymo's large-scale driverless safety analysis, Waymo's disclosure on remote assistance, Waymo's funding and ride-volume update, Waymo's description of its driving and validation system, and Waymo's work on large-scale autonomous-driving simulation.
Commercial and deployment evidence also comes from Pony.ai's latest quarterly results, Pony.ai's Gen-7 operating update, and Zoox's public robotaxi launch in Las Vegas. For scenario-based testing, we used ISO 34505.
Regulatory and safety context comes from California DMV's autonomous-vehicle regulations, California's updated safety-case and first-responder requirements, California's AV permit information, NHTSA's automated-driving crash-reporting program, and NHTSA's vehicle cybersecurity guidance.
Additional reference points include the review of autonomous vehicles and first-responder interactions, AAA's survey on public trust in self-driving vehicles, California DMV's Cruise suspension, General Motors' decision to stop funding Cruise's robotaxi development, and McKinsey's autonomous-fleet cost estimates.
This chart, included in our autonomous vehicle market deck, illustrates yearly VC funding for autonomous vehicle startups
Related blog posts
- Autonomous Vehicles: what are the biggest challenges now?
- Autonomous Vehicles: what is getting real adoption now?
- Autonomous Vehicles: what is actually working now?
- The latest funding news in the autonomous vehicle market
