Domain randomization is a powerful technique #Arduino #robotics #engineering #education #AI #ML

本站添加: 2026-05-25

[00:00:00]The little bot is much more robust now.

[00:00:02]You can see I gave it a pretty solid push and it will try to come back into place and settle into one of those grooves. Now, it oscillates back and forth a little bit and that's because I could not simulate the ridges in the tire. MuJoCo did not like that. So, I had to randomize those using a technique called domain randomization. Speaking of that, I randomized a bunch of stuff and did a lot more curriculum learning.

[00:00:27]So, after I did initial training in phase two that you saw last time in the previous post, I randomized things like readings from the sensors, how long it takes to actually execute an action which is up to one time step.

[00:00:41]In phase four, I randomized the amount of motor noise.

[00:00:48]I gave it little random pushes.

[00:00:51]I also randomized the mass and friction.

[00:00:56]How much the motor could actually drive the wheels and that's due to things like voltage sag.

[00:01:02]So, that's what I'm trying to simulate here. So, running off a battery versus being plugged in or if the voltage sags when you try to run the motors.

[00:01:11]And then, here's where I tried to simulate those ridges in the tires because MuJoCo did not like that.

[00:01:18]I simply just created random torques on the axles, little bitty torques to kind of simulate that and it sort of learns it. You can see it actually training here.

[00:01:29]It's just kind of sitting there. Oh, there is an episode reset. And then, you can see how it performed over each of those phases here and each time I introduced a new randomization, you can see it mostly forgot what it was doing and then kind of got back up to fully balancing for the entirety of the episode. Now, to accomplish that, I also increased the size of the neural network by using 32 nodes per layer, and that seemed to work much better. And now it's off to balancing.