You must have read tones of hyping articles about autonomous cars over the last 5 years. Yet, they are still nowhere to be found.
Why is it so exactly?
Making autonomous car is a hard task, but we have solved plenty of hard tasks. For instance, we have improved computers to go from occupying one full room to occupying your jean pocket.
How did we manage to do so? By constant, small progresses.
And that’s autonomous car problem: it’s all or nothing.
First let’s recap why making an autonomous car is really hard. If you think about it, an autonomous car has two fundamental blocs.
- One is about perception, that is sense and make sense of what is around itself. The sensing part is already done, although with high cost sensors. The hard part is making sense of everything. A sensor is going to tell that there is a small white object on the road. But what is it really? Markings? Snow? plastic bag?
Don’t forget that you are not only trying to understand objects, but also human behavior of other people behind the wheels! That’s a tall order. - The second building block is making decisions. Even with a perfect knowledge and understanding of your environment, you still need to decide what to do every second you drive.
Assuming you have a perfect map, and a perfect idea of where you are on that map (localization), your work is still very hard, because now you have to account for the other road users. Worse, you have to decide what to do now, and what future consequences it will bring.
Now you are hopefully convinced that building autonomous cars is hard. And no gradual improvements either, we have to take a all-or-nothing approach because of a basic geometry problem.
You see, we use standard levels to measure #autonomouscar 🤖.
🚗 At level 2 : the car drives, but you have your hands on the wheel, and you monitor everything all the time
🚗 At level 3, you can watch a movie, but still, you handle emergencies
🚗 At level 4 and 5, the car can fully drive without you.
I don’t know about you, but to me L2 is of little interest, as you still have to focus on the road, so you’re still driving! L4 and L5 are perfect but super hard to reach technologically. On the surface, L3 could seem like a good compromise: you only drive for emergencies 🚨.
What a perfect blueprint for researchers: we could just start at L2 and make small improvements until we reach full autonomy.
Sorry, but nop ! That’s where the tyranny of turning radius comes.
Let’s take a typical car traveling at typical speed ~90km/h (25m/s) on a typical turn.
At this speed, you can see only 4 seconds ahead of you, and that is just pure geometry, so this constraint is not going anywhere. Now a typical car would need about 2 seconds to brake to a full stop.That leaves only 2 seconds
- A. the car warns you
- B. you understand what to do
- C. you react
Ok, let’s go over these 3 steps.
A. is not too bad, you car could give visual and sound clues in ~0.1s.
C. could also be OK. If you know what to do, you can react in the ~0.3s, like a sprinter at the start of the race.
The real problem is understanding what to do (B.) ! Remember the last time you were very focused on your movie, and somebody asked you something? Remember it took a couple of seconds to snap out of the movie and answer that question? Well, sadly it takes us human up to 10 seconds.
So we can’t have L3 autonomous cars, we have to jump directly from L2 to L4 and L5. Hence the hype (because autonomous cars are super cool), and the wait (and misery for us researchers).
— — — — — — — — — — — — -
Hypothesis : road is 16m wide (including 8m of roadway), turning radius of 80m based on official specs, speed of 90km/h (25m/s). Harsh braking.
