End-to-end sensor fusion for localization: Visual-Inertial-Wheel Odometry

During my thesis at the Mechatronics Group at Aalto University, supervised by the Deep Learning professor at Eurecom, I had implemented and extended a deep learning pipeline presented in this paper. The goal is to estimate the vehicle localization by fusing features extracted from different sources of odometry in a fully end-to-end fashion, meaning that no parameters need to be tuned when the system is assembled, whilst making the system more robust to sensors failure or disturbances.

​

Answering the question "Where am I?", while learning tools that can be leveraged in a much broader field (such as the Deep Learning tools) , has been a fascinating research topic. I will keep working on this topic as a Visitor Researcher at Aalto University, to dive down into the world of representation learning to understand how robots can learn the task of localization, maybe taking the biological world as inspiration.