Can you confirm you are using both MotionGC (gyroscope calibration) and MotionGT (gyroscope temperature effect calibration)?
Keep in mind that those libraries are opportunistic meaning that when they get a chance, they will use the data from the sensor to calibrate and compensate.
For MotionGC this means that the library is looking for steady periods, where the gyro bias can be reliably computed by averaging gyro output samples. If there are no steady periods, then this will never happen. Note that gyro misconfiguration may block MotionGC because high level of noise may be interpreted as motion. Vibrations may have a similar effect. In your application, there must be periods where gyro is steady and there are no vibrations so that MotionGC can do its work.
Similarly for MotionGT: the library is looking to learn the gyro bias at different temperatures, and will try to model the hysteresis (if there is any). If the temperatures at which the bias is learned are clumped together because the device is in a temperature-controlled environment, then MotionGT can do nothing when temperature goes out of the learned range. For example, if you 'calibrate' at room temperature of 20-25 Celsius (70-80F), then you go outside in a colder/hotter environment 5 Celsius or 35 Celsius (40F or 100F), the library will not be able to compensate anything.
Gyro bias is subject to changes at every power toggle, and because of many other factors such as mechanical stress on the board, or temperature gradients. If any of these factors are present, the bias compensation may not be as effective as it is expected to be.
