Adaptive Lane Following
Design of a controller to autonomously tuning the kinematics parameters of an AV while it while it autonomously roams around a city.
The project was developed and tested on the Duckietown platform, in the scope of the Autonomous Mobility On Demand: From car to fleet course at ETH.
This works addresses the problem of uncertainty in knowing the exact values of the measures from which the kinematics of the vehicle. While the common procedure involves performing a manual wheel calibration before starting using a Duckiebot, the solution developed here by me and my teammate Simone Arreghini, allows to avoid this bothersome step, making the AV adjust its knowledge of the kinematics parameters while roaming around the city performing lane following.
Even though this system builds up on the assumption that the Duckiebot starts with a reasonable assumption over the values of these parameters, intensive testing proved that, apart from very borderline cases, the proposed solution visibly improves the stability of the vehicle over the course of its roaming.
The images below show the response of an originally uncalibrated Duckiebot to a “go straight” command, before and after the on-the-run calibration for 15 seconds.
If you are interested in knowing more, the source code relative to this project and a brief report are available at the following link: |
