My guess is that the ODOM_TWIST_COVARANCE2 matrix is corrupted, and the Y variance should be 1e-9 to match the X variance, with no non-zero, off-diagonal entries.