Available Theses
Interested in our work? Come work with us! Here is a non-comprehensive list of the topics for M.Sc. theses. Details will be defined depending on the candidate's background and objectives. Each thesis work may lead to scientific publication. Contact me for more info.
AI-based scene understanding for autonomous systems
This thesis explores how semantic segmentation, monocular depth estimation and object detection models can be used to obtain a detailed map that can be employed to build a digital twin. This method can be applied either for autonomous driving, railway systems, or other.
Prerequisites: python (PyTorch)
Est. time: 6 months
Exploring noise models for LiDAR sensors
LiDAR sensors have become crucial for scene perception and understanding. Simulators easily recreate synthetic point-clouds by using raytracing, but synthetic sensor data lack real-world degradation effects typical of the atmospheric conditions, such as rain and fog. The objective of this thesis is to devise and validate realistic noise models that can be used to improve the realism of the point-clouds generated with simulators.
Prerequisites: C++/python
Est. time: 6 months
Development of an optimized framework for Unreal Engine 5 - ROS2 communication
This thesis will provide a custom solution to allow Unreal Engine to connect to ROS2 and exchange messages in real-time. The objective is to design and implement an UE5 plugin that is able to publish and receive messages on ROS2 topics with minimum latency also for large-scale data streams such as images and point-clouds.
Prerequisites: C++, knowledge of Unreal Engine and/or ROS2 is recommended
Est. time: 6 months
Exploring the Simplex architecture: recovering from AI errors
This thesis will investigate how to detect when a neural network (RL agents, or perception modules) is making a mistake in the context of autonomous systems. The objective is to design a "Safety monitor" module that will devise whether the neural network provides wrong outputs (e.g., when adversarial attacks are ongoing or an out-of-distribution sample is detected); in that, case the Simplex architecture will switch to a safe, certifiable controller. The code will be developed in simulation (Gazebo, UE5) and then deployed on a real-world rover.
Prerequisites: deep learning/RL, python/C++, knowledge of ROS2 is recommended
Est. time: 6 months
Beyond the empty room: from virtual to augmented reality
Our current setup simulates and bypasses the camera and LiDAR inputs to the rover. The rover’s perception of the real world (which in our case is an empty room) is limited to LiDAR data to avoid collisions against the walls. This project focuses on the extension to the case of “mixed reality”. The rover is now placed not in an empty room, but in a safe environment (e.g., an empty parking lot), where we want the rover to perceive the real world, but also include additional virtual objects in the camera and LiDAR data. The objective is to devise techniques and algorithms to superimpose virtual objects in the real-world perception of the rover.
Prerequisites: C++, knowledge of UE5 is recommended.
Est. time: 6 months
Testing adversarial attacks in the real world
Adversarial attacks are still an underdetermined threat for deep learning perception systems. Real-world adversarial examples might be dangerous for AI-based autonomous systems. This project focuses on the evaluation of the effects of such attacks for autonomous systems, especially against camera-based perception. The objective is to devise strategies and algorithms to detect whether an attack is undergoing, exclude the corrupted command, and design countermeasures to avoid collisions. The evaluation will be performed at different reality levels (fully simulated, vehicle-in-the-loop simulation, testing in the real world).
Prerequisites: deep learning, python (pyTorch recommended), ROS2 recommended.
Est. time: 6 months