Why it’s so hard to make a reliable self-driving car

Two decades ago, the US military kicked off the race to build self-driving cars by sending a fleet of fledgling robot vehicles across the Mojave Desert in its seminal Darpa challenge.

By 2015, autonomous vehicle technology was being widely pursued, and the industry was abuzz with predictions that driverless cars would soon be everywhere. 

It did not happen. Some legacy carmakers and start-ups abandoned their efforts, citing excessive costs and complexity. Regulators stepped up scrutiny of the emerging technology after crashes involving cars equipped with partial-automation systems. 

The companies still making progress are proceeding with caution, aware of the heavy reputational damage when someone is injured or killed by a self-driving car.

Waymo,

a venture of Google parent Alphabet

, introduced its driverless taxi service to a fourth major US city in March through a

partnership with Uber Technologies

.

Tesla

launched a “robotaxi” service

with roughly 10 or 20 Model Y cars operating under human supervision in parts of Austin, Texas, in June. It was still a far cry from what chief executive Elon Musk pledged in 2019, when he said the carmaker was a year away from putting one million robotaxis on the road. 

Who is leading the way in driverless cars?

Mr Musk eventually wants to offer a fully autonomous ride-hailing service using a purpose-built two-seater called Cybercab. Production of the vehicle – a prototype of which has no steering wheel – may start in 2026.

But Tesla has a track record of unmet promises, and it still is not clear how the company intends to make the leap from selling advanced driver-assistance features to full autonomy.

Federal regulators are investigating the suite of driver-assist features that Tesla markets as Full Self-Driving after multiple crashes, one of which resulted in a fatality. The National Highway Traffic Safety Administration said it was looking into incidents where Tesla robotaxis appeared to violate traffic laws during their first day offering paid rides in Austin. 

Waymo is making the biggest strides for now, with its robotaxis well-established in Phoenix, Los Angeles and San Francisco. The company plans to bring them to Atlanta this summer after debuting its service in Austin in March.

Waymo is looking to incorporate some of its technology into Toyota’s passenger cars, which could democratise autonomous-driving features. 

China has been a hotbed of activity, thanks to dozens of start-ups, with several companies expanding globally.

WeRide, which in May secured another US$100 million (S$128.6 million) investment from its US partner Uber, plans to expand its robotaxi offering to Dubai and Europe.

Baidu is planning to test and eventually launch its Apollo Go autonomous ride-hailing service in Europe. It has one of China’s largest robotaxi fleets, operating in cities such as Beijing and Guangzhou.

At times, however, it can seem as if every advance is matched by a serious setback. 

Who has been having problems?

Ford Motor and Volkswagen shut down their self-driving business Argo AI in October 2022, a dramatic turn of events for a business that had more than 2,000 workers and for a time considered an initial public share offering.

Ford, which wrote down its US$2.7 billion investment in Argo, formed a new unit to focus on driver-assist features.

VW’s chief executive Oliver Blume later axed Audi’s plans for a self-driving vehicle because of slow progress. The German carmaker is still making some efforts, including via partnerships with suppliers Robert Bosch, Mobileye Global and Horizon Robotics.

Apple sank billions of dollars into trying to develop a vehicle with so-called Level 4 or 5 driving automation capability before deciding in February 2024 to shut down its car project. Several smaller driverless-car start-ups have abandoned their efforts. 

General Motors ended its Cruise venture in 2025 after one of its robotaxis struck a pedestrian in San Francisco in October 2023. The company shifted resources towards trying to develop autonomous vehicles it could sell to consumers and improving its Super Cruise driver-assistance features.

In May, GM hired Mr Sterling Anderson, a former head of Tesla’s Autopilot efforts. 

Aurora Innovation announced in May that it was putting a human back behind the wheel of driverless trucks in Texas, reversing course less than three weeks after starting a commercial service there. Aurora said the decision to move an “observer” from the rear of the cabin into the driver’s seat was made at the request of truckmaker Paccar.

What is the challenge with full autonomy?

To get a passenger safely to his or her destination, Level 5 autonomous systems must be able to deal with any conceivable situation in every possible location, in all weather, day or night. All this requires a lot of sensors and sophisticated computers that gobble up computing power and are expensive to develop. 

Some of the unlikely scenarios and freak incidents they will need to cope with may not even have happened yet. Split-second manoeuvres can involve decisions that require complex trade-offs that can be a challenge even for a human. For example, do you swerve and collide with an elderly couple to avoid killing a child? 

Even if self-driven cars end up being involved in fewer accidents than those driven by humans, regulators have less tolerance for mistakes made by a machine, partly as the technology’s proponents have touted improved safety as one of the advantages of vehicle autonomy. 

What are the alternatives to full autonomy?

Advanced driver-assistance systems – known as ADAS – are helping drivers to park, stay in their lane and avoid objects using cameras, radar and other sensors. The systems can alert drivers and, in some cases, briefly take control of the car to avoid collisions. 

Carmakers are increasingly offering ADAS in their vehicles. One of the first lower-level automation offerings was the anti-lock braking system, introduced more than four decades ago, which became standard equipment. More recent systems offer features such as emergency braking and automated parking.

How smart are these systems?

The industry categorises automation systems from Levels 0 to 5. Level 0 features simply pass on information to the driver, like sounding a warning when he or she is driving out of a traffic lane. 

Tesla’s Autopilot is classified as Level 2 because it requires constant driver input and supervision, much in the way a pilot oversees certain automated systems in a plane’s cockpit. 

Mercedes-Benz is offering Level 3 automation – requiring neither hands on the steering wheel nor eyes on the road – in select vehicles under certain conditions in parts of Germany and the US.

The robotaxis being tested in the US and China could be categorised as more advanced Level 4 systems, but these vehicles are limited in terms of where they can go. The pinnacle – which has yet to be achieved – would be Level 5 cars that can drive autonomously everywhere and in all conditions. BLOOMBERG