-
Notifications
You must be signed in to change notification settings - Fork 508
ekf: disable xy accel bias learning before takeoff #818
Conversation
As those biases are usually poorly observable before takeoff because they are almost perpendicular to the gravity vector, learning is often driven by noise and numerical issues. This results in incorrect bias learning before takeoff when the drone is static on ground for a long period of time.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It would be good to have the same log with good velocity or position observations.
CI tests are checking if the variances are big enough after some initialisation time. These values are set based on previous results. They can be changed without hesitation. |
Can you do a test flight on uneven ground? Then we could replay before and after and see the difference. |
Should be sufficient tolerance given acos of 0.8 is ~36 deg. |
@bresch It would be beneficial to also add unit tests for this. |
@kamilritz Add the checks that the correct axes are not learned in the roll90 and pitch90 initialization? Would that be enough or do you have another check in mind? |
Yes, for all main axis pointing down, a check that only the right bias is learned |
replaces #816
follows #817
In the log below, I oriented sequentially each axis with the gravity vector
Note that the accel bias learning is quite bad because the only sources of aiding are baro and fake hpos.