Simultaneous State Estimation and Tire Model Learning for Autonomous Vehicle Applications

This article addresses the problem of state estimation and simultaneous learning of the vehicle's tire model on autonomous vehicles. The problem is motivated by the fact that lateral distance measurements are typically available on modern vehicles while tire models are difficult to identify and also vary with time. Tire forces are modeled in the estimator using a neural network in which no
a priori
assumptions on the type of model need to be made. A neuro-adaptive observer that provides asymptotically stable estimation of the state vector and of the neural network weights is developed. The developed observer is evaluated using both MATLAB simulations with a low-order model as well as with an unknown high-order model in the commercial software CarSim. Cornering and lane change maneuvers are used to learn the tire model over an adequately large range of slip angles. Performance with the low-order vehicle model is excellent with near-perfect estimation of states as well as the tire force nonlinear characteristics. The performance with the unknown high-order CarSim model is also found to be good with the tire model being estimated correctly over the range of slip angles excited by the executed vehicle maneuvers. The developed technology can enable a new approach to obtaining tire models that are otherwise difficult to identify in practice and depend on empirical characterizations.