Thesis

We offer a wide variety of thesis opportunities for students. Available topics include Algorithmic Fairness, Reinforcement Learning, Smart Mobility, Anomaly Detection, Continual Learning, Visual Anomaly Detection, Object Recognition (YOLO), Weakly Supervised Learning, Active Learning, Explainability, and Predictive Maintenance.

For Reinforcement Learning–based theses, students may choose to carry out their work either in collaboration with a company or at the university. In both cases, a PhD student will be assigned as a mentor to provide guidance, suggestions, and support in case of major difficulties.

Students are also welcome to propose their own thesis topics of interest. Any self-proposed topic must be discussed with the professors to ensure feasibility and alignment with the group’s research directions.

Some thesis proposals are:

The cost of learning dilemma

In the last years, Artificial Intelligence (AI) has become an increasingly central actor in our lives, enabling technological solutions to adapt to the immediate needs of the end-users. The shift from a static to a continuously evolving technology does not come for free but presents a critical cost in terms of communication bandwidth, computational power, and other network resources. Hence, telecommunication networks are facing a competition between traditional data flows and those supporting the training of AI algorithms. This leads to a new dilemma: how can we balance network resources among end users and learning agents?

The scientific community ha starting being aware of the issues associated with the communication and computational overhead due to AI. Despite this, most works still assume that learning and user data are exchanged on separate channels, ignoring the dependency between the agent training and network conditions. To overcome such a limit, it is necessary to investigate how to implement AI solutions from a new perspective.

This project aims at defining new strategies for combining the optimization of network resources and learning algorithms, analzying how the improvement of the AI training implies a resource reduction for target applications. The project consists in identifying significant use cases where the "cost of learning" dilemma arises, modeling the scenarios theoretically, and proposing new solutions for the described problem.

For more information, contact Federico Mason at federico.mason@unipd.it
Edge Computing-based Automated Driving in Duckietown
Duckietown is a simple platform for automated driving, in which small cars need to navigate a town with pedestrians, traffic lights and signs, and other vehicles. The cars use an onboard nVidia Jetson board to run the computer vision and decision-making algorithms, but there is a reasonable need for offloading this processing to a central node. In this case, the wireless channel will play a significant role, as the latency deadlines are very tight and the data that need to be transferred can have a significant size.
In this thesis project, you will work to evaluate and improve deep reinforcement learning-based driving algorithms in constrained communication scenarios in Duckietown. Aside from the standard simulator, a physical testbed is available in the lab, and working with it will be a major component of the project.
PREREQUISITES: in order to work on this topic, you should have completed the "Neural networks" and "Reinforcement learning" courses.
CONTACTS: federico.chiariotti@unipd.it
Security in Goal-Oriented Communications
Goal-oriented or effective communication is a new paradigm that allows sensors to only transmit the data that is relevant to the receiver, e.g., by omitting redundant or irrelevant information. However, this poses significant security challenges: if the transmitter uses dynamic compression, adapting the length of packets to the required amount of data, an eavesdropper may glean some information about the system from the timing and length of packets, even if their content is protected through encryption. This project involves the design of a learning-based eavesdropping attack on an effective communication system, evaluating the performance of the attacker and the possible countermeasures that the transmitter can employ.
PREREQUISITES: in order to work on this topic, you should have completed the "Neural networks" and "Reinforcement learning" courses.
CONTACTS: federico.chiariotti@unipd.it
Deep Reinforcement Learning for sensor scheduling
Wake-up radio is a technology that allows Internet of Things (IoT) nodes to respond to requests, which can be ID- or content-based (in the former case, the sensor will send its latest reading if its ID is in the request, while in the latter, it will transmit if the data matches the conditions specified in the request message). This type of system allows for interesting scheduling opportunities, particularly if the sensor measurements are correlated: by carefully crafting scheduling requests, the IoT gateway can save energy and improve its estimate of the state of the system.

This project involves the design of a Deep Reinforcement Learning algorithm to achieve these two goals at the same time by using knowledge on the system state and statistics as well as the sensors' battery states and capabilities. The problem can also include other factors, such as selecting different objective functions over the system state.
PREREQUISITES: in order to work on this topic, you should have completed the "Neural networks" and "Reinforcement learning" courses.
CONTACTS: federico.chiariotti@unipd.it
Goal-Oriented Integrated Sensing and Communication
The Integrated Sensing and Communication (ISAC) paradigm uses wireless communication to sense the environment, using multi-antenna systems to infer, e.g., the position, identity, and current activity of people in a room by learning the perturbations they cause in a wireless channel. This is a very promising approach for 6G, as it would allow the network to be aware of what happens in its surroundings, adapting to events in real-time without the need for external sensors and cameras. However, the communication burden from transmitting all the channel sensing data for processing and learning is significant. Effective communication is a smart compression technique that can allow nodes to abstract the relevant information in the data, transmitting only the features that are necessary for the task at hand, e.g., for identifying the person by their gait and physique. The thesis project involves analyzing ISAC data and exploiting deep learning techniques to perform semantic compression, reducing the communication footprint of the target application while maintaining the same performance.
PREREQUISITES: in order to work on this topic, you should have completed the "Neural networks" and "Reinforcement learning" courses.
CONTACTS: federico.chiariotti@unipd.it
Reinforcement Learning for Durable Algorithmic Recourse
Algorithmic recourse provides individuals rejected by an automated decision system with counterfactual explanations for the reason behind their rejection. These explanations are often interpreted as recommendations to improve the likelihood of future acceptance. Previous work has leveraged Reinforcement Learning to ensure the robustness of such recommendations in competitive, resource-limited settings (https://arxiv.org/abs/2509.22102). However, additional layers of complexity and contextual constraints can be incorporated into the modeling. In particular, while prior work accounts for endogenous shifts in the reapplying population, one should also consider that new candidates’ scores may be influenced by these recommendations. Furthermore, to ensure equity, it is important to prevent recommended scores from oscillating excessively across time steps. In this thesis, we extend the simulation to include these modeling elements and adapt the reinforcement learning solution to remain robust under such complexities.
PREREQUISITES: in order to work on this topic, you should have completed the "Reinforcement learning" course.
CONTACTS: marina.ceccon@phd.unipd.it
Title: Reinforcement Learning for Energy-Efficient Cloud Robotics

Abstract: Cloud robotics is a computing paradigm where compute-constrained robotic platforms (e.g., lightweight UAVs or mobile robots) offload compute tasks to a cloud server with virtually unconstrained computational power [1,2,3,4]. This lets robots benefit from sophisticated data processing, such as inference of large machine learning models (e.g, DNN for semantic segmentation), while saving on local compute and battery. The other side of the coin is that transmitting data from robots to the cloud and viceversa can severely congest the network, leading to unacceptable delays for time-critical tasks such as autonomous driving. Therefore, wisely choosing when to offload compute jobs and when to perform data processing locally on robots to improve system performance without incurring large delays is paramount.

This thesis explores the delicate tradeoff discussed above. You will be tasked to learn an effective RL-based policy for compute allocation while considering performance (e.g., classification or detection accuracy), network latency, and limited robot battery. In particular, you will consider one (or multiple) low-compute, low-accuracy inference models run locally on the robot and a high-accuracy model run at the cloud for image classification and semantic segmentation. However, if you'd like to explore different tasks, they can be discussed with the supervisor before starting the thesis.

This thesis may involve a – possibly remote – collaboration with the Massachusetts Institute of Technology.

[1] Chinchali et al., "Network Offloading Policies for Cloud Robotics: a Learning-based Approach," Autonomous Robots, 2021. Available at https://arxiv.org/pdf/1902.05703

[2] Hu et al., "Cloud Robotics: Architecture, Challenges and Applications," IEEE Network, 2012. Available at https://personal.ntu.edu.sg/wptay/MyPapers/Journals/HuTayWen%20-%20Cloud%20robotics%20architectures%20challenges%20and%20applications.pdf

[3] Nakanoya et al. "Co-design of communication and machine inference for cloud robotics," Autonomous Robots, 2023. Available at https://www.roboticsproceedings.org/rss17/p046.pdf

[4] Liu et al. "RoboEC2: A Novel Cloud Robotic System With Dynamic Network Offloading Assisted by Amazon EC2," IEEE Transactions on Automation Science and Engineering, 2024. Available at https://ieeexplore.ieee.org/abstract/document/10347007
Title: Multi-Agent Reinforcement Learning for Compute Allocation in Processing Networks

Abstract: Processing networks are multi-agent systems where each agent is a smart sensor equipped with sensing, compute, and communication capabilities. Agents acquire information from the environment, refine it via local compute (e.g., machine learning or computer vision), and transmit data to a workstation that implements global monitoring and coordinated decision-making for all robots. This setup is common in Internet-of-Things and Edge Computing, where smart sensors may represent collaborative robots for manufacturing and the workstation is a controller that commands robots' motion. Since smart sensors carry limited resources, they face a tradeoff between latency and accuracy when processing sensory measurements onboard: either they extract accurate information with extra delay, or they process data quickly at the cost of moderate extracted information. Clearly, balancing this tradeoff is key to success and high performance in time-critical applications.

In this thesis, you will use multi-agent reinforcement learning to find an efficient policy that decides how each smart sensor in a processing network refines – if at all – sensory data before transmitting them to the workstation, drawing inspiration from and extending the frameworks in [1,2]. The primary focus will be optimal global estimation/monitoring with use cases from robotic mapping and coverage. If progress and time permit, we may move from monitoring to decision-making. In any case, use cases will be agreed with the supervisor before starting the thesis to make sure they match your interests.

This thesis may involve a – possibly remote – collaboration with Purdue University.

[1] L. Ballotta, G. Peserico, F. Zanini, and P. Dini, “To Compute or Not to Compute? Adaptive Smart Sensing in Resource-Constrained Edge Computing,” IEEE Trans. Netw. Sci. Eng., 2024. Available at https://arxiv.org/abs/2209.02166

[2] V. Tripathi, L. Ballotta, L. Carlone, and E. Modiano, “Computation and Communication Co-Design for Real-Time Monitoring and Control in Multi-Agent Systems,” International Symposium on Modeling and Optimization in Mobile, Ad hoc, and Wireless Networks, 2021. Available at https://arxiv.org/abs/2108.03122
Title: Graph Neural Network-Based Safe Multi-Robot Control

Abstract: Multi-robot systems operate under stringent safety requirements such as avoiding collisions and respecting mutual distances. In recent years, control barrier functions (CBFs) have emerged as a gold standard to satisfy such constraints in real time, by ensuring set invariance at all times (i.e., the robots' state remains inside a safe set) with a small computation overhead. One limitation of CBF is that it is usually difficult to design it, which often leads to conservative choices that reduce performance.

In this thesis, you will explore a learning-based approach to improve the performance of CBF for multi-robot systems, focusing on collision avoidance between robots [1]. If time and progress permit, nonideal elements such as sensor noise and delays will be factored in to make the problem more realistic. If desired, simulations can be performed with ROS and Gazebo.

This thesis may involve a – possibly remote – collaboration with the National University of Singapore.

[1] L. Ballotta and R. Talak, “Safe Distributed Control of Multi-Robot Systems With Communication Delays,” IEEE Transactions on Vehicular Technology, 2025. Available at https://arxiv.org/abs/2402.09382
Title: Resilient Distributed Federated Learning via Trust Observations

Abstract: Distributed federated learning is an emerging distributed computing paradigm where agents collaboratively learn machine-learning models via fully distributed (peer-to-peer) communication. Agents in the network do not need to share their own data, but exchange only model weights and/or gradients with neighbors, retaining privacy, and aggregate such received information with local models to improve accuracy. However, a potential issue of this setup is that malicious agents can covertly intrude the network and pollute the training, to reduce accuracy of models learned by the agents. In particular, if malicious agents are not promptly detected, they can irreversibly disrupt the collaborative learning process and ultimately degrade monitoring or decision-making tasks of normal agents.

You will be tasked to devise a resilient aggregation strategy that is robust to unknown adversaries, namely, it does not overly degrade accuracy of models learned by normal agents. A starting point, to be discussed with the supervisor, is the robust aggregation algorithm in [1] that integrates opinion dynamics to shield normal agents. Also, we will consider using trust observations that normal agents obtain from the wireless transmission channel to assess the legitimacy of transmissions up to a certain confidence [2,3].

[1] L. Ballotta, N. Bastianello, R: M. G. Ferrari, and K. H. Johansson, “Personalized and Resilient Distributed Learning Through Opinion Dynamics,” 2025 [under review at IEEE Transactions on Control of Network Systems]. Available at https://arxiv.org/abs/2505.14081

[2] M. Yemini, A. Nedić, A. J. Goldsmith, and S. Gil, “Characterizing Trust and Resilience in Distributed Consensus for Cyberphysical Systems,” IEEE Transactions on Robotics, 2022. Available at https://arxiv.org/abs/2103.05464

[3] L. Ballotta, Á. Vékássy, S. Gil, M. Yemini, “Confidence Boosts Trust-Based Resilience in Cooperative Multi-Robot Systems,” 2025 [under review at IEEE Transactions on Automatic Control]. Available at https://arxiv.org/abs/2506.08807