Why Tesla Won’t Use LIDAR ? And which technology is ideal for self-driving cars
Almost every single company working on self-driving cars right now uses LIDAR. Uber, Waymo, and Toyota all use it, but not Tesla. I want to go over what the two competing technologies have to offer and what we should expect from self-driving cars in the future.
Lidar VS Vision
Lidar is a method of measuring distance by shooting lasers and detecting how much time they take to return. The idea is similar to Radar but instead of radio waves, we use lasers. The technology is extremely accurate at detecting objects even up to millimeters.
Computer Vision is a field of Artificial Intelligence that trains computers to understand the visual world. This is basically reverse engineering human vision.
Tesla’s Vision
Tesla has been heavily relying on Vision and going against LIDAR sensors. At the same time, all the other companies use Lidar and do not seem to care. Elon Musk even said:
LIDAR is a fool’s errand… and anyone relying on LIDAR is doomed. — Elon Musk
If you would like to see all Elon’s thoughts on the technology choice, make sure to check his talk at Tesla Autonomy Day.
Cost
The most apparent reason for Tesla to take a different route is the cost. The cost of placing a single LIDAR device on a car is somewhere around $10,000. Google with its Waymo project has been able to slightly decrease the number by introducing mass production. However, the cost is still rather significant.
Tesla is highly focused on costs and making sure the cars are affordable. Adding the prices of a LIDAR on top of the already expensive car is quite significant.
Application to real roads
One of the most important point and the one emphasized highly in the conference is the correlation with human vision. As humans, we do not throw lasers in every direction to be able to drive a car. Neither should the cars, as mentioned by Elon Musk.
Everything we see on the road is full of visual information. All the signs, turns, intersections are there to help us navigate. All of these are stationary objects and it is great that LIDAR is so accurate in detecting them.
Issues start to arise when moving objects appear on the road. Humans, dogs, flying plastic bags, are all objects we frequently encounter on the road. LIDAR is not able to detect how they are moving or even what those objects are.
LIDAR cannot differentiate a road bump from a plastic bag. This is example was mentioned in the conference and is extremely important to take into consideration. If we are driving at high speeds on a highway and there is a plastic bag, we do not need to make a quick stop. It will not be much of an issue if we hit it.
Now, if the car stops, that’s where real dangers come in. Cars behind will probably not be able to react so quickly to our stop in the middle of the road. Such situations further demonstrate the attention to detail required when making self-driving cars.
Tesla made it clear that their system of cameras and radar is able to detect what an object is. The radar looking forward is able to quickly tell if there is anything problematic ahead. Once an object comes into sight, cameras will decide what the object is and then the car can react to the situation.
Adaptation
Another essential takeaway from Autonomy Day and other interviews with Elon Musk is that the system is made to adapt. They talked a lot about the Neural Network used and how the system is able to use the data provided to make rational decisions.
One of the big issues with Tesla’s competitors is the lack of such adaptiveness. Most of these systems either rely heavily on high-accuracy maps with road lines or were never tested on real roads. Yes, we have seen Waymo drive around cities. However, only the large roads with highly efficient maps. The lighting, weather conditions, and traffic are ideal in those demos. Most people do not drive on such roads every single day.
Smaller roads, with unexpected turns and lanes that change in size, are quite a bit more common. Plus, Tesla is an actual car that you can buy. People have driven over a billion miles on Tesla cars, while Waymo was only tested on around ten million miles.
The amount of difficult and unpredictable road data that Tesla was able to accumulate is invaluable. That is how the system learns and keeps getting better. Such a concept is actually quite promising since customers actually see constant improvements.
Accuracy
Another interesting point can be deduced from a research paper published by Cornell University. The paper discussed how stereos cameras can be used to generate a 3D map nearly as accurate as a LIDAR map.
We can conclude from the paper that other than spending $7,500 on a LIDAR device, you could get a few cameras that only cost $5 and get nearly the same accuracy. So when the guys at Tesla say that such hardware will become outdated soon, they could have a point.
Final Thoughts
With how much money is poured into the field of self-driving cars and the constant competition, we can be quite optimistic that such cars are coming. Whether Tesla will be the company that does this, we cannot know and might not even need to. Actually, there could potentially be several ways of developing self-driving cars. We might even end up seeing the combination of the two, which would not be that surprising.
When it comes to Tesla, we could actually purchase such a car. They are driven all over the world and the progress could actually be observed in real-time. You cannot buy a Waymo or an Uber car.
The reasons for Tesla’s decision were covered here, however it does not mean we should all bet on Tesla. Maybe some other company manages to create a self-driving vehicle first. What we do know, is that some company will do it and we should pay close attention to everything happening in the field.
