SLAM with Reinforcement Learning to aid mapping in highly dynamic scenes

This project aims to develop a reinforcement learning agent with access to depth cameras to aid map building for SLAM. Research has shown dynamic environments, where the majority of frame is taken up by dynamic objects, along with low fidelity and lighting can lead to low robustness in practical application. We hope to use reinforcement learning as a motion planner to explore the environment to better enable static landmark detection and dynamic landmark filtering. This work will first validate the hypothesis through simulation. More specifically using the Unreal Engine with AirSim and a pre-modeled physical quadrotor and flight controller. This simulation will explore how reinforcement learning can be used to navigate through a prior unknown environment to aid dynamic-SLAM to detect static landmarks. Then a real-world experiment will be setup within the drone arena along with dynamic objects to evaluate the robustness in the real world. We will be building upon previous work that has been done within simulation (https://www.youtube.com/watch?v=LL5iKaSM2VQ) along with using a quadrotor that has already been built within the lab (https://www.youtube.com/watch?v=bmtHIAjggP8).

Faculty Supervisor:

Sajad Saeedi

Student:

Partner:

University of Bath

Discipline:

Computer science

Sector:

Artificial Intelligence; Technology; Transportation (excluding aerospace)

University:

Toronto Metropolitan University

Program: