A Unified Perturbative Dynamics Approach to Online Vehicle Model Identification

The motions of wheeled mobile robots are largely governed by contact forces between the wheels and the terrain. Inasmuch as future wheel-terrain interactions are unpredictable and unobservable, high performance autonomous vehicles must ultimately learn the terrain by feel and extrapolate, just as humans do. We present an approach to the automatic calibration of dynamic models of arbitrary wheeled mobile robots on arbitrary terrain. Inputs beyond our control (disturbances) are assumed to be responsible for observed differences between what the vehicle was initially predicted to do and what it was subsequently observed to do. In departure from much previous work, and in order to directly support adaptive and predictive controllers, we concentrate on the problem of predicting candidate trajectories rather than measuring the current slip. The approach linearizes the nominal vehicle model and then calibrates the perturbative dynamics to explain the observed prediction residuals. Both systematic and stochastic disturbances are used, and we model these disturbances as functions over the terrain, the velocities, and the applied inertial and gravitational forces. In this way, we produce a model which can be used to predict behavior across all of state space for arbitrary terrain geometry. Results demonstrate that the approach converges quickly and produces marked improvements in the prediction of trajectories for multiple vehicle classes throughout the performance envelope of the platform, including during aggressive maneuvering.