This paper describes an online method to calibrate certain kinematic parameters of legged robots, including leg
lengths, that can be difficult to measure offline due to dynamic deformation effects and rolling contacts. A kinematic model of the
robot’s legs that depends on these parameters is used, along with measurements from joint encoders, foot contact sensors, and an inertial measurement unit (IMU) to predict the robot’s body velocity. This predicted velocity is then compared to another velocity measurement from, for example, a camera or motion capture system,
and the difference between them is used to compute an update on the kinematic parameters. The method can be incorporated into both
Kalman filter or sliding-window optimization-based state estimator. We provide a theoretical observability analysis of our method,
as well as validation both in simulation and on hardware. Hardware experiments demonstrate that online kinematic calibration can significantly reduce position drift when relying on odometry.