The recent YouTube video from Mark Rober’s channel was quite interesting. The title was “Can you fool a self driving car”, which is very catchy as all the YouTube videos do. The video is actually an introduction to LiDAR technology.
The first half of the video was about mapping the famous Disneyland’s dark ride, Space Mountain, using LiDAR. If you know how LiDAR works, you would understand this is technically possible. But what is important is that Mark Rober actually did it. He even reviewed the park’s guidelines and found that they don’t prohibit using LiDAR devices in the park. I’d bet they will add that rule soon.
Fooling the car is in the second half of the video. It was comparing Tesla self driving, which depending on solely vision data, and another autopilot technology which adopted LiDAR. The comparison is especially focused on extreme conditions, like very heavy rain, dense fog and a painted road in a wall. However it turns out the video appeals the benefits from LiDAR. Although it was very informative, it was also a framing play.
One thing you should not miss while watching this video is the difference between “driving assistance” and “self driving”. The tests in this video were the features of a driving assistance rather than that of a self driving. It is sure that a driving assistance technology helps making a better self driving technology. But it is not the only feature of a self driving technology.
I think that I understand why Tesla decided to depend on vision data only for their self driving system. And I agree on their decision. The input for the learning process of a self driving system is “what does a human see in a given driving scenario” and “how does a human respond to it”. And LiDAR is not a part of what a human sees while driving. So in this context, the LiDAR data might be considered noise. Even after watching this video, I still believe in this approach.
However, I believe an additional “driving assistance” with LiDAR can be helpful even for Tesla self driving system. This video shows that there are cases that cameras cannot see but LiDAR can. Then the challenge would be the cost of adopting LiDAR to Tesla again and the benefits from it. If the benefit outweighs the cost, I believe Tesla would consider it. If not, they won’t. But in the end… who knows?
Kirin's Technology Essay 