2021-06-01 20:47:48
In this post, I explain a complex self-driving algorithm in few words. If you enjoy such posts please like it to give me feedback. Otherwise, you can just view the nice video about self-driving racing cars. Thanks!
AV system should find the right control signals to move a vehicle in the right direction. There are a lot of problems making the task complex. First of all relation between control signals and vehicle movement is not linear. A second one, there are a lot of possible trajectories with different qualities. Next, there are other agents and objects on the road. Some of them have unpredictable behavior. And some unexpected events can radically change vehicle trajectory: some stone under a wheel, wet surface, and even wind.
There are a lot of algorithms to solve the problem. One of them is called Model Predictive Path Integral Control (MPPI). Initially, it was developed to control AutoRally vehicles model with one-fifth scale. Look at the video. I especially make the link on the moment where you can see problems: inertion, unexpected road grip, and stones.
Without details algorithm is simple. Let's assume that at each moment we have some control sequence for the next few seconds. Initially, it can be a "do nothing" sequence. At each step of the algorithm, we create a few thousand new control sequences. Then each sequence is converted to trajectory using a vehicle dynamic model (based on a neural network). Each trajectory is evaluated by multiple parameters, e.g. drive time, sharpness, obstacle crossing, etc. Finally, the best trajectory is selected and a first control step from it is executed.
By repeating the algorithm at 40Hz you can drive the car as shown in the video. Sampled trajectories are shown as fans ahead of the car.
And as usual, you are welcome to give me feedback in private chat. The channel still in beta and I'm looking for the best format.
Video:
Article: https://arxiv.org/pdf/1707.04540.pdf
67 viewsedited 17:47
D
