Камери Tesla Autopilot не помічають перешкод / YouTube
Tesla Autopilot collided with a painted road on a fake wall located in the middle of the road. This happened during a test comparing cameras and lidar.
Most companies developing autonomous driving technologies use a combination of different sensors (cameras, radar, lidar, and ultrasonic sensors). However, Tesla insists on using only cameras. The automaker has removed radars from its models and even disabled those that were already installed in cars.
So far, this strategy has not yielded the expected results, as Tesla’s systems remain only at Level 2 in the autonomous driving classification, which means that the driver must constantly monitor the movement.
The company’s CEO Elon Musk says that Tesla’s main advantage is the ability to quickly scale the technology once the problem of autonomous driving is solved. According to him, the vision system based on cameras and neural networks works like a human driver, and therefore will be able to adapt to any road.
Critics disagree with this statement, especially after Musk said that Tesla will achieve Level 5 autonomy, which means full self-driving in all conditions. Cameras have limitations that can be solved with the help of lidars, so his words are questionable.
A new video from engineer and YouTuber Mark Robert vividly demonstrated this problem.
In the video, Robert tested a Tesla Model Y with autopilot compared to a car equipped with a lidar system in various conditions.
The Tesla with Autopilot was able to stop in front of a child’s dummy in the middle of the road when it was standing still, moving, or illuminated by bright headlights. However, the car did not stop in fog or heavy rain.
Not surprisingly, lidar, a system based on lasers, is better at recognizing objects in thick fog than cameras.
The result in heavy rain was more unexpected, but it is worth noting that the downpour in the test was very intense.
The last experiment with a fake road painted on the wall in the Wile E. Coyote style was obviously unrealistic. But it illustrates well the difference between cameras and radars or lidars: cameras estimate possible obstacles based on an image, while lidars collect precise data about objects.
Simply put, lidar doesn’t pay attention to what’s painted on the wall – it just recognizes the wall itself, whereas cameras can be fooled. In fact, Tesla Autopilot coped with only half of the test tasks.
So far, it seems that in the world of Elon Musk and Tesla, the technology «I see a person» is more like «I can be fooled». After all, a person can recognize a road painted on a wall, but Tesla cannot. Perhaps it’s time to make changes to the design of cars and install more modern and sophisticated sensors, as other manufacturers do? We can only hope that in the future, Tesla Autopilot will stop mistaking scenery for reality and will recognize children in bad weather.
It’s worth noting that Tesla’s autonomous driving systems don’t perform well in real-world road conditions either. Recently Cybertruck with FSD v13 crashes into a pole. At the same time Tesla drivers are massively fined for using FSD because the system does not adequately assess the situation on the road.
Source: Electrek