Revision history [back]

The package isn't doing any explicit logic to infer the velocities; the correlation between velocity and pose means that, when we generate the Kalman gain, we end up getting velocity values, even if only the pose changed. The velocity prediction in the kinematic model is just current velocity + acceleration * dt, so once your odom messages stop being published, you're just going to see the filter move on with a constant velocity until you get another input from position or velocity.

Other than monkeying with the process noise and measurement covariances (the latter appear to not be set), I wouldn't expect to have a lot of control over that behavior.