In an attempt to understand the issue of the robot constantly moving away from its start position, data from previous tests was plotted.

This first plot showed that the sensor thought the angle was drifting away from equilibrium; this explained why the robot kept moving in the same direction and the slope downwards explained the acceleration. Next it was decided to see how the PID error value was changing:

The above plot revealed more evidence for gyro drift. We can see the system error is desired (mean of approx 0) yet the sesnor thinks the angle of the robot is slowly decreasing over time; this is causing movement acceleration away from the starting position to keep itself "balanced". Next it was decided to test the MPU6050 readings outide of a balancing test:

Three tests were conducted each of about 3 minutes duration. 1) sensor was left undisturbed, 2) sensor was given a quick shake then left undistrubed, 3) sensor was randomly disturbed for 3 minutes by hand then placed level again. The results verified the gyro dirft and also showed that the drift is worsened by constant excitation.

Ther current dirft compensation strategy is to determine a set of calibration constants that are determined using an average drift over 500 loops calculated during the robots start up procedure, these constants are then subtracted from the gyros raw values for every program loop. The above results show that this method is not working very well. Two options were considered: 1) calculating the average drift over more loops (1000 loops perhaps) 2) deploying a more intelligent (dynamic) drift compensation technique that could recognise false readings and adapt its influence accordingly.

It would be great to get in touch with someone who has worked with gyro drift in the past.