Calibration Best Practises

I would be grateful if knowledgably folks could verify my understanding gleaned from the docs and MotionApps/Drive code.

I'm working on producing rock-solid calibration for our MPU-6050 based device. We will do this post-assembly but also offer it in our host app so end-user may perform it if necassary.

So;

All MPU6050's are factory calibrated. There are some 3rd party libs out the wild that enable read/write of these non-volatile values. I'm not sure why this is needed.

The sensor has a DMP-based gyro calibration mode. When enabled DMP will calculate gyro bias after 8 seconds of no motion and write these values into DMP memory (D_EXT_GYRO_BIAS_X/Y/Z). On startup, with the sesnsor not moving, the DMP outputs can take 15 seconds or so to stabilise (8 s of which is the 'no motion' window).

The new gyro bias values are volatile but can be queried and persisted by the host (e.g. in EEPROM) and then pushed to DMP on startup (dmp_set_gyro_bias). This would reduce the settling time (with DMP Gyro Calibration disabled).

Accelerometer calibration can also be performed by the client code and pushed to DMP (dmp_set_accel_bias). Again these can be calculated by the host (or queried via 'self test') and stored in EEPROM to be pushed to DMP on startup.

Accelerometer calibration would ideally be done in zero-g :) But since we're stuck on Earth we need to account for gravity. By placing the sensor flat we ensure only one axis (z) experiences g. So bias for z we have to subtract g in order to take it out of the equation. We DONT want the final Z bias value to bias out gravity.

The mpu_run_self_test function in inv_mpu.c runs the built in self-test routines and returns new bias values for gyro and accel based on raw gyro and accel data. If self-test has passed (which it should) then the new bias values can be scaled and pushed down to DMP (as described above).

In summary;

mpu_run_self_test provides fresh calibration bias values that the host can persist in EEPROM and push to DMP on startup.

There is no need to mess with factory bias values.

Cheers,

Rob J.