In March, Jensen Huang, chief executive of computer chip giant NVIDIA, declared the “ChatGPT moment” for self-driving cars had arrived. In Australia, Tesla’s Full Self-Driving (FSD) system is already available on public roads. Waymo is exploring robotaxi operations in Australia.
The question is no longer whether autonomous driving will arrive, but whether Australia is ready to operate it safely, efficiently, and sustainably. This raises an infrastructure question: can roads, bridges, and intersections designed for human drivers also be understood by machines?
Over the past 100-plus days, my colleagues and I have used a Tesla Model Y with FSD every day on Queensland roads. We recorded more than 500 safety-critical events where the system required driver intervention or revealed an important limitation in how it interpreted the road environment.
To make some of these observations useful beyond our own research, we initiated White Box Autonomy as a public archive of events where autonomous vehicles hit problems in real-world conditions.
Two things stood out. First, the technology is already far more capable than many people realize. In many cases, the vehicle handled the driving task smoothly, with a precision that is hard for a human driver to match.
Second, it is also less capable than many people assume. It occasionally struggled with situations human drivers handle almost automatically, making mistakes that were surprising and sometimes dangerous because they weren’t the kinds of errors an experienced driver would make.
Where the system struggled
There is a small bridge in my neighborhood that I had barely noticed before. Like most drivers, I simply drove over it without thinking too much.
But almost every time Tesla FSD approached it, the vehicle appeared confused and began weaving from side to side.
This pointed to a broader issue: road markings and layouts that are easy for humans to interpret can still confuse autonomous driving systems.
Time-restricted speed limits, such as school zones, caused similar problems — we had to intervene more than 90 percent of the time. In one case, FSD obeyed a school-zone limit correctly, except it was evening, long after school hours.
Railway crossings and boom gates were another concern. In one incident, the car ahead stopped near a railway crossing. Had FSD continued moving, it could have ended up stopped on the tracks. I had to slam on the brakes.
Australia’s zipper merge rule — simple in principle, but reliant on subtle human judgment and informal negotiation — also proved challenging. In one incident, neither FSD nor the other vehicle slowed down at the merge point, and I had to intervene quickly.
There were other recurring issues. FSD often struggled at complex roundabouts (Australia has no shortage of them), misread poorly marked or steep streets crowded with parked cars, mistook e-scooter riders for pedestrians despite their very different behavior, and lost accuracy in harsh or variable weather that obscured lane markings and road edges.
In more than 100 days of daily testing, we did not complete a single trip in which FSD drove independently from start to finish without intervention.
Some of this represents problems with the technology. But some could be addressed by adapting the roads it operates on.
Smarter vehicles need smarter roads
The original vision of autonomous driving looked very different from today’s. In the 1990s, programs such as California PATH imagined vehicles supported by smart roads and controlled highway environments.
Today, the logic has reversed. The focus is on making vehicles smart enough to navigate existing roads unaided.
But smarter vehicles alone aren’t enough. The real opportunity is to meet in the middle: improve vehicle intelligence while also making infrastructure more friendly for autonomous vehicles.
This doesn’t mean rebuilding everything or filling streets with expensive technology. Many fixes are basic: clearer lane markings, more consistent signs, better-maintained road surfaces, less ambiguous intersection design, and reliable speed-limit information.
For example, instead of a single speed-limit sign or a bicycle symbol painted only at the start and end of a lane, agencies could repeat these cues more often, giving automated systems more opportunities to recognize them, correct misreadings, and adjust safely.
AV can also help infrastructure
Autonomous vehicles shouldn’t only be seen as users of the road network — they can act as mobile sensors within it. As they travel, they can flag potholes, faded markings, damaged signs, and bottlenecks, helping agencies detect problems earlier and prioritize maintenance.
Real-time data from autonomous vehicles could also support congestion mitigation, incident detection, and traffic management. Autonomous vehicles can be part of the system that monitors and improves the network, not just a user of it.
Fix the common problems first
Researchers often focus on “edge cases” — rare, unpredictable situations that are hard for automated systems to handle. These matter, and remote assistance may help vehicles manage them.
But our testing suggests a more immediate priority: most of the incidents we saw weren’t rare edge cases, but everyday situations. If Australia wants to be ready for autonomous driving, governments and the automotive industry should work together on these repeatable problems by making infrastructure more friendly for autonomous vehicles.
Many of these changes can be made in the process of ordinary public works and maintenance — but new infrastructure and major events such as Brisbane’s upcoming 2032 hosting of the Olympic and Paralympic Games present opportunities for easy, lasting improvements.
Zuduo Zheng, Professor, School of Civil Engineering, The University of Queensland
This article is republished from The Conversation under a Creative Commons license. Read the original article.
Follow us on X, Facebook, or Pinterest
