We found a small systematic error in our calibration procedure for the robots which led to a fixed offset between actual and observed joint positions. This was fixed and all robots have been re-calibrated.
This is effective for all submissions being executed after 01.12.2020 15:10 CET.
You can fix the values in old logs by subtracting the offset from the robot observations:
offset = numpy.array([-0.02433333, -0.043, -0.00366667, -0.01833333, 0.029, 0.01633333, 0.04266667, 0.014, -0.01266667]) fixed_position = observation.position - offset
Likewise, in case you trained a policy that already adapted to this error and want to restore the old behaviour, you should be able to do this by adding the above offsets to your observations. Note that when using position actions the offset has to be considered there as well.
I am very sorry for this issue and hope it does not cause too much trouble.