Saturday, March 7, 2009

Assembling the self-balancing robot

Well, Alan and I both got tired of waiting for me to finish the modeling and simulation I wanted to do, so I threw caution to the winds and ordered all the parts to build my draft design of the self-balancing robot. There is still the chance that the motors may not have sufficient speed or torque, or may have too much backlash to work well. Also, I went ahead and ordered the cheapest accelerometer available at Sparkfun, (the MMA7260) reasoning that we’d have a go at making it work first. We can always switch to a different sensor or add additional sensors if we have to.

So now Alan is spending this weekend assembling the robot. I have written Rev 1 of the software, and soon we will be putting them together for our first trial runs. For Rev 1, we are not going to try to balance. We have put a small caster on the front of the box so it can drive around as a typical three wheel vehicle. Once we debug the motor control and get a feel for the speed and torque we have available, I will be back trying to develop the balancing control software.

In the meantime I have been testing the accelerometer and, true to everyone’s comments, the measurements are alarmingly noisy. Of course the signal can be filtered, but that introduces delay in the measurements. Whether we can get this to work or not will depend, I think, in large measure on how rapidly the main control loop has to run to keep the robot well balanced. The MMA7260 has a refresh frequency on the X and Y axes of 350 Hz. So you get a new reading about every 3 milliseconds. If we can afford to make corrections to the motor inputs only every 30 ms, say, then we can take the average of the last 10 readings to smooth out the acceleration data. If this turns out to be too slow, the robot will be unsteady and we’ll see lots of random jittering back and forth. On the other hand, noisy acceleration data will cause jittering, too. So we must find the best tradeoff. But I note with some unease that there is still significant noise in the data even after averaging over 10 readings. And, of course, the more heavily the signal is filtered the more delay it creates before the control algorithm will see the beginning of an excursion. This can cause oscillation or even a loss of control. I’ll try to post some specific data at some point so you can see what I’m talking about.

Alan and I have had a number of discussions about whether we needed wheel encoders so we can get feedback on the distance traveled by the robot. We are going to build it first without encoders. This, too, is heresy among self-balancing robot builders (uh, I mean builders of self-balancing robots). Partly the standard wisdom arises from how people conceptualize the inverted pendulum problem. The natural way to think about it is that I measure how far the pendulum has departed from vertical, then I move the base that far to get it back under the center of gravity. But of course, for balancing, you get all the position feedback you need from the pendulum itself. However, the more fundamental reason people assume you need encoders is that the accelerometers can’t tell you whether you’re moving, only whether you’re accelerating. There is no way to correct the inevitable error you get from computing velocity by integrating acceleration. So when the robot is commanded to stand still it has no way to be sure it really is, and when it is commanded to move at a certain speed it has no way to determine that either.

My solution to this problem is partly electrical and partly anthropic, if you will. One piece of data we do have is the average current being supplied to the two motors in the forward and backward directions. To some approximation, you expect that if the robot is sitting still on a level surface the average power in each direction will be equal. Of course, there are natural physical irregularities so that is not exactly correct. But I believe we can add a bias term to that balance and treat it as a tuning parameter. We can adjust it to correct for (almost all of?) the drift. That’s the electrical part. The other part is based on the idea that this is not actually an autonomous vehicle. It is just a fancy remote control car. So the human in the loop will be controlling the position and speed to their satisfaction. Of course, this approach does have another downside: the robot will not be able to maintain position on a slope by itself. It will slowly roll downhill. But I think this will seem like a very natural behavior to the human operator, and of course they will compensate.

The only real question, which we’ll find out by building the thing, is how well the drift can be corrected with a constant bias, that is, how stable the bias is. Just in case, though, I bought some optical encoders and eventually I expect we’ll get around to fooling with them – if not on this vehicle, then on the next.

We’ll have another report, with pictures, when Alan finishes construction. I can’t wait!

2 comments:

Rob said...

There must be a better way to clean up the accelerometer data than using a simple average. Maybe you can keep a running estimate of what the robot "thinks" the accelerometer should read based on its recent motor movements and such. You could combine that estimate with the instantaneous data from the accelerometer using a Kalman filter to improve your estimate of the current acceleration.

Of course, all this takes processor time and space, so you'd have to be careful about the complexity of the acceleration model. I suspect you need a physics model like this to drive the wheels anyway. The only new thing to add is the update scheme.

Its also uber-cool to be using mathematics that is commonly (?) referred to as "linear quadratic estimation"!

Bill Hensley said...

Yes! If you recall from my first post a few weeks ago, I am going to try to get away without using a gyro, unlike most designs. One of the ways I hope to compensate is by basing the control strategy on a a physics model. So I want to do something like you suggest, but I don't know exactly how to go about it. That's why I'm going to need the help of my control engineer friends. (Hint, hint...) At work I think I have Tom sufficiently intrigued to help, and also perhaps Hong.