Autoevolution has covered The Dawn Project’s recreation of Mark Rober’s test showing that Tesla Vision cannot detect a wall which is painted like the road. 

This article was published in Autoevolution on 28 March 2025, and can be found here.

A few weeks ago, a famous YouTuber showed that Teskla Vision cannot tell reality from a painting. The initiative featured a Luminar prototype vehicle and was harshly criticized as biased. Others have recreated the video, including prominent Tesla Critic Dan O’Dowd, with puzzling results. The key takeaway is that HW3 vehicles consistently fail the test, dashing hopes that they could become fully autonomous.

If I had $100 for every time Elon Musk said Tesla is on the verge of cracking vehicle autonomy, I’d probably be able to pay for the FSD package. However, that would still not make a Tesla EV fully autonomous, as the EV maker is not yet ready to assume responsibility for the software’s actions. Instead, it officially refers to the FSD and Autopilot as Level-2 driver assistance systems, to the disappointment of Tesla fans.

With the latest FSD iterations, Tesla hit “a fork in the road,” to use one of Musk’s favorite expressions. On the one hand, older cars fitted with HW3 Autopilot computers have become a bottleneck for FSD development. On the other, even the HW4 computer might prove limited in achieving Tesla’s goal. Elon Musk has been cautious about this, saying that HW3 might not be powerful enough to make a car drive itself.

HW3 limitations have become more obvious in light of a recent PR stunt by famous YouTuber Mark Rober. He compared Tesla’s Vison-only FSD system with one that also includes LiDAR sensors. To prove his point, Rober installed a painted wall that looked like the continuation of the road and drove both cars into it. As predicted, Tesla’s Vision system failed to discern between reality and the painted wall, whereas the LiDAR-equipped car correctly detected the obstacle and braked the car.

However, many inconsistencies in Rober’s video made it look like a Luminar promo material. Soon after he published the controversial video, Kyle Paul, a Tesla enthusiast, recreated the test using both an HW3 Tesla Model Y and an HW4 Cybertruck. Like Rober’s car, the former failed to detect the wall, but the Cybertruck stopped. However, many people thought the Cybertruck saw the wall because the lighting made it obvious against the surrounding landscape.

Paul is working on a new test to eliminate the lighting inconsistencies, and we’ll probably see the results soon. Until then, there’s another person pitting HW3 and HW4 cars against each other in the wall test. Dan O’Dowd, a long-time fan and, at the same time, a fierce critic of Tesla and Elon Musk, has recreated the already famous “Looney Tunes” test.

Surprisingly, many people considered O’Dowd’s test an honest attempt at showing Tesla Vision’s limitations. His tests showed that HW3-equipped vehicles running FSD V12.6.4 cannot detect the wall. However, his subsequent tests show that HW4’s improved sensor suite is insufficient to perform safely. When driving on Enhanced Autopilot, even the HW4 car failed to detect the wall. Only when using FSD V13.2.8 did the car brake for the painted wall.

This appears to show that hardware alone is not enough, and the results vastly depend on the software’s capabilities. It’s unclear whether Tesla could fine-tune FSD on HW3 cars to detect such obstacles or if it is only possible for newer hardware. If only the latter is true, Tesla will have to spend a lot of money retrofitting older vehicles to HW4 or even HW5 if Tesla moves further by then.

Musk promised to offer free upgrades of HW3 vehicles for owners who bought FSD. A majority of Tesla EVs sold in North America are HW3, and in some markets, the EV maker has installed HW3 computers until mid-2024. This might make retrofitting all these vehicles a costly business.