Adaptive Curvature Control with a Radial Basis Function Network for Autonomous Driving

Master thesis defense

Philipp Ebner

Monday, December 06, 2021, 11:00


This thesis presents an adaptive control method for the curvature control of an autonomous car. The currently used cascaded control architecture with an outer-loop position controller to generate the target trajectory remains unchanged, while we focus to improve the inner-loop curvature controller. The proposed control scheme combines approximate model inversion (AMI) with a direct adaptive controller similar to model-reference adaptive control (MRAC) and a radial basis function network (RBFN). The dynamical curvature model used in AMI is derived from the linear single track model. Based on the curvature control error, the parameters of the RBFN are adapted online. The derivation of the adaptation law for the parameters of the RBFN is built on Lyapunov stability theory. Furthermore, we present an optimization-based random track generator, which provides sufficiently different maneuvers to test the algorithm. Since the RBFN is linear in the parameters, the computational effort required for the resulting adaptive controller is low. Experimental results in simulation with CarMaker and on a real car show improved tracking performance of the curvature controller compared to the currently used model-based controller.


Fully online.
Closed to the public, but open team internally.
We will use the syscop group meeting internal zoom
(contact for the link).

For legal reasons, attendance will be documented.