Model-based control techniques are widely used for autonomous navigation, but achieving good performance in the presence of disturbances and uncertainties requires online estimation and adaptation of the model. Agricultural applications are a typical example where the variability of the environment and the complexity of the wheel/terrain interaction model require an approach based on a simple model of the robot/vehicle, equipped with algorithms capable of continuously adapting/improving this model using all sizes available.

This thesis aims to develop classical and machine/deep learning techniques to estimate online the slip of a track, constituting the fundamental element for the creation of a controller for the autonomous navigation of a skid-steering robot for offroad/agricultural applications.
Different sensors will be taken into consideration, starting from the state of the robot and the information provided by an IMU, up to the images of the terrain generated by an RGBD camera or a lidar.
A key aspect of the thesis will be the comparison between classical and learning-based techniques, in order to discover any added value of the latter.
The developed algorithms will be validated in simulation and/or in the field using a tracked robot.

This work is suitable for both students coming from an engineering and non-engineering background, and can be tackled both by a single student and by a group of two students. Knowledge of Python is required.

Reference teacher: Prof. Luca Bascetta – luca.bascetta@polimi.it