Self-driving vehicles were supposed to be everywhere by now.
In a December 2015 interview with Fortune, Elon Musk said Tesla was about two years away from producing a fully autonomous vehicle that could operate “in any condition and on any road.”
That timeline passed without the technology reaching that level.
More than a decade later, however, there is at least some semblance of the future Musk described. Autonomous vehicles, or AVs, can operate without human drivers in several cities, and commercial services have expanded most visibly in China and the US.
Baidu’s Apollo Go provided 3.2 million fully driverless rides in the first quarter of 2026. By April, it had passed 22 million public rides, and by May, it was operating in 27 cities. Its fleets had accumulated more than 220 million fully driverless kilometers.
In the US, Waymo has similarly expanded its commercial robotaxi operations. Its latest safety analysis covered more than 220 million fully autonomous miles.
But these services are still some way from the definition Musk laid out in 2015.
The gap between Level 4 and Level 5
The Society of Automotive Engineers’ (SAE) framework categorizes driving automation across six levels, from Level 0 to Level 5, based on the degree of autonomy a vehicle can provide.
Level 4 represents high driving automation. Vehicles can handle the entire driving task without human intervention, but only within defined areas and conditions. Level 5 represents full driving automation, where a vehicle can perform the entire driving task without human intervention regardless of the road or traffic conditions.
There are currently no commercial Level 5 AV services operating at scale. Most fully driverless commercial services today, including those operated by Waymo and Baidu, fall under Level 4.
That raises two questions:
- What makes the jump from Level 4 to Level 5 so difficult?
- And does autonomous driving even need to reach Level 5 before it can play a meaningful role in transportation?
At its simplest, moving from Level 4 to Level 5 means removing the operating boundaries that current autonomous vehicles depend on. A Level 5 vehicle would need to handle roads and situations it has not been specifically prepared for, rather than operating within a predefined area or set of conditions.
That is a major technical challenge because current autonomous driving systems can rely partly on highly detailed maps of the areas in which they operate, corroborating that information in real time with data captured by cameras, LiDAR (light detection and ranging), or other sensors. Waymo, for example, describes using detailed maps alongside its onboard sensors to help the vehicle navigate its service areas.
Removing those geographical constraints places more responsibility on what the vehicle can perceive and interpret on the road at any given moment. It must be able to respond safely to unfamiliar layouts, unusual human behavior, temporary road changes, and other edge cases without relying on the same level of prior knowledge.
Level 4 is already proving useful
That does not mean autonomous vehicles need to reach Level 5 before they become useful.
Level 4 systems are already being deployed for specific transportation needs. Beyond Baidu and Waymo, AV developers including Pony.ai and WeRide have explored robotaxis and autonomous shuttles in different configurations.
Singapore offers one example. The country has been working on AVs for more than a decade. The Land Transport Authority (LTA) said it has been testing whether they could be integrated into the land transport network since 2014, with one-north becoming an on-road testbed in 2016.
That work is now visible in Punggol, where three fixed-route autonomous shuttle services are being progressively deployed for first- and last-mile connections. Grab operates two routes using WeRide’s technology, while ComfortDelGro operates the third using Pony.ai’s.
Community rides began in January this year, while public rides on two Grab routes began in April. By July, more than 11,500 people had taken the services.
When KrASIA tested one of the services, the shuttle was driving itself. A safety operator remained in the driver’s seat, however, and the vehicle followed a fixed route, stopping only at designated pickup and dropoff points.
The limitations are clear, but so is the utility. An autonomous shuttle does not need to travel anywhere, under any condition, to connect residents with transport nodes or fill gaps in an existing public transportation network.
There is also evidence that autonomous vehicles can outperform human drivers on some measures of road safety.
Waymo’s analysis of more than 220 million fully autonomous miles found 94% fewer crashes causing serious or fatal injuries and 82% fewer injury-causing crashes compared with human drivers in the same areas.
Singapore’s early passenger feedback has also been positive. In LTA’s July post-ride survey, nearly 99% of around 900 Punggol shuttle users said they felt safe and would recommend the service.
An incident on Edgedale Plains in January offers another illustration of how the technology can behave differently from human expectations. During a road test, a ComfortDelGro AV detected an object that was not actually on the road and moved into an adjacent lane. The safety operator, seeing no obvious reason for the move, took control. The vehicle then hit a road divider. LTA’s subsequent simulation found that the AV would likely have completed the maneuver safely if the operator had not intervened.
Statistical safety is not the same as trust
But evidence of safer performance may not necessarily persuade passengers to use AVs.
There is a less tangible difference between a vehicle being statistically safer and passengers feeling that it is safe.
People may be less comfortable when an AV makes a decision they do not understand or would not expect from a human driver. Not everyone has acclimatized to the idea of sitting in a vehicle without someone at the wheel.
That may partly explain why, when KrASIA tried one of the Punggol services, a safety operator remained in the driver’s seat throughout the journey, with their hands close to the steering wheel, even though the vehicle was capable of navigating the designated route autonomously.
Singapore also requires AVs to meet technical performance and public acceptance metrics before they can carry passengers and eventually operate without a safety operator. LTA said these requirements include completing sufficient distance on the actual deployment route without intervention and handling different traffic situations within the authorized area.
Before passengers are carried, the shuttles undergo closed-circuit testing and on-road preparation without passengers. Routes are mapped, sensors are calibrated, artificial intelligence models are validated, and safety operators familiarize themselves with the routes and local traffic conditions.
These safeguards can help make a defined Level 4 service viable. But they also demonstrate why scaling the same system across an entire country is considerably harder.
Convenience matters too
Suitability is another constraint.
The value of an autonomous vehicle depends on the transportation needs of the region in which it operates. In many places, deploying Level 4 systems remains difficult because of infrastructure requirements and the technical work needed to prepare vehicles to operate reliably within a particular environment.
The limitations of current services can also make them less convenient than existing alternatives.
A vehicle may appear superior on paper because it is safer and does not require a human to operate it. But if it can travel only along preset routes or within predetermined areas, it may not offer the same mobility as driving a private vehicle or taking another form of transportation.
Punggol highlights this tradeoff. In the same survey where 99% of respondents reportedly said they felt safe and would recommend the shuttle, almost 60% wanted more flexibility over where they could board and alight, while around 40% wanted more direct routes.
LTA has since been working with Grab on an on-demand model within the existing network, giving passengers more flexibility over where they board and alight.
A service, in other words, can be safe and driverless but still lose out if another option is more convenient.
Deployment depends on local conditions
Above the technology, safety, and convenience questions sits regulation.
Authorities have to decide how autonomous vehicles, regardless of their level of automation, can coexist with the rest of a transportation system.
Singapore would appear to offer relatively favorable conditions. Its road networks are mature, it has spent years testing AV technology, and private vehicle ownership is expensive, creating an incentive to explore other transportation options.
Even then, deployment has taken years.
LTA has been exploring AV integration since 2014. In 2022, the Ministry of Transport (MOT) said wider deployment would depend on AV technology meeting safety standards and gaining public acceptance, and that the technology was not yet ready for large-scale, citywide deployment.
In May 2026, MOT said it was developing a broader legal framework for AVs covering safety, accountability, liability, insurance, and enforcement. Singapore’s existing road traffic framework was designed around human drivers.
Authorities may also have to consider disruption to existing transportation services. Robotaxis, for example, could compete with human taxi and private-hire drivers if they become a reliable and cost-effective alternative. A 2026 study published in online journal Humanities and Social Sciences Communications found that the introduction of Apollo Go’s robotaxis in Wuhan was associated with a 10.9% short-run decline in traditional taxi drivers’ average daily income.
The equation can look very different elsewhere. In economies such as Indonesia and Vietnam, where personal motorcycles play an entrenched role in everyday transportation, the conditions for introducing autonomous vehicles at scale may be substantially different.
This means there may be no single timetable for AV adoption. The pace will depend not only on what the technology can do, but on whether a city has the infrastructure, regulations, transportation needs, and public acceptance needed to make deployment worthwhile.
Level 5 may not be the point
Musk’s 2015 prediction was about a car that could drive itself “in any condition and on any road.”
More than a decade later, the industry is still not there if Level 5 is the north star. Level 4 vehicles can drive themselves, but only within boundaries.
Yet those boundaries do not necessarily make the technology useless. Waymo and Baidu are already using Level 4 systems commercially. Singapore is testing autonomous shuttles as part of its public transportation network. Other operators are exploring similar models around the world.
Rather than fixate on whether a vehicle can drive itself everywhere, attention may be better focused on where autonomous driving already works well enough to be useful but still lacks the right conditions for deployment.