Dynamics based Observer for Differential-Drive Vehicles

Localization, i.e., pose estimation of a ground vehicle with respect to some inertial frame, is among key issues in mobile robotics. Majority of approaches employed to solve this problem depend primarily on fusion of information from multiple sensors. Frequently, these approaches rely on kinematics model of vehicle for predicting forward in time during fusion. The kinematic models are simple and fast, but do not account for dynamics. The use of dynamics based fusion framework can lead to relatively accurate prediction by exploiting information from the cause of the motion, which kinematic model based frameworks fail to do. Additionally, when depending mainly on exteroceptive sensors such as GPS, Vision, Lidar, etc, for localizing, there may be cases where the sensors may fail temporarily due to changes in environment which may cause localization accuracy to suffer degradation or even loss of localization. The information from dynamics can come in handy for localizing in such scenarios. In this work, we derive a continuous-time dynamic model, which also takes friction into account. Additionally, we propose a Continuous-Discrete Extended Kalman Filter(CD-EKF) based observer, which uses the derived dynamic model as process model for prediction, and measurements from position sensor as measurement updates. Moreover, we perform experiments to show that proposed observer can auto-calibrate the dynamic model in real time, which is more versatile than methods which calibrate dynamic models beforehand, as it enables adaptation to changing conditions like a terrain transition, payload or tyre air-pressure etc.