Comment by Thomas for Dear Okasa-san,as for mobile robots, every frames on top of base_link are actively relying on the localization / SLAM node and therefore the result you will obtain will depend in these particular cases will depend on how it behaves.In my case, I use robot_pose_ekf (on HRP2) so that I both combine the integration of the command and the IMU data. The odom is published by this node like on PR2.For base_footprint, I use the point defined as the center of the robot soles center. If you walk on non-flat floor, of course you have to project this point at the right height depending on your perception. This is not an issue on flat-floor although. This definition works fine for any number of legs.If the robot is falling then of course the "footprint" is not really a footprint anymore but it is still defined (mathematically speaking). As for the odom frame, by changing the covariance matrices appropriately you can only integrate the IMU data when you fall to get an estimate of your position.P.S. FYI the robot_pose_ekf node input is an odometry message sent by the robot controller (hrpsys in my case).

The `base_footprint` rotational part is undefined in the current state of the REP. IMHO this is not the best frame to express contacts so I would keep the rotational part to identity if I were you. That does not prevent you to add additional tf frames to express contact points position.