[Tom Bottiglieri]

We were using a similar setup to 121 for version 1 of our traction control code. The only sensory was 4 encoders (2 on driven wheels, 2 on "follower" wheels). I am pretty sure some of the more controllable systems have logic to throttle on current spikes, but I digress... The logic was basically to hold slip to 0 on each wheel when the robot is stopped, allow a little slip on launch, and allow a bunch of slip on turns.

Before you get into tuning the controller, I would analyze your feedback. If the resolution isn't good enough, you will end up with some noise in slip ratio between cycles. This effect will be made even worse if you are taking percent slip (Vw-Vf/Vw), as you are pushing something big (encoder velocity) into something small (dividing by a velocity), then blowing it up (multiplying by Kp). The end result of this will be uncontrollable "bucking" by your drive wheels. Some nested filters (one on slip ratio, one on motor output) should be able to fix this, but this will involve just another level of tuning.

The method we used to tune our controller was to take off from a stop, and adjust the P gain until acceleration was optimized. We used a set length and timers to do this. We didn't have enough time to optimize I, but I would probably drive the robot into a large weight and adjust the I gain enough so it can slow down the wheels enough to push it, without overshooting or "bucking". The I term in this controller, as I understand it, exists for situations just like these where an unexpected external force is applied to the base. Your I gain should be very small in comparison to the P gain. (I want to say about 10^3ish smaller, but I could be wrong)

Good luck.