Academic

This section collects some of the my academic work, and especially projects developed for my MSc exams. They are reported in reverse order, from latest to oldest.

As mentioned in the home page, I hold a BSc in Computer & Systems Engineering and I’m currently studying for a MSc in Control Engineering. The courses I chose mostly focus on robotics, with some machine learning and non-robotics applications of control theory.

MSc #

Pick and Place blockstacking task resolution using Graph Neural Networks (Neural Networks / October 2026) #

In this project I replicate results from two different papers (linked in the materials). The objective is to perform pick and place tasks with a robot manipulator, using a Graph Neural Network (GNN) to encode the scene and predict which block to pick and where to place it. In particular I focus on blockstacking tasks: starting with a tower of blocks in some initial position and random order, the objective is to invert the order of the tower at another position. The network shows good generalization to scenes with more blocks that it was trained on, but does not generaliza to blocks that have different sizes, possibly indicating a data normalization problem. Some of my initial experiments also show that good results can be obtained by using a directed graph whose edges encode the relationships between blocks and goals, instead of a fully connected, undirected graph as in the two reference papers.

Repository ○ Materials (Presentation with videos + jupyter notebook + model checkpoints)

Neural MPC for Control of Type 1 Diabetes (Control of Communication and Energy Networks / July 2026) #

In this project we build a grey-box model mixing a LTI model and a LSTM neural network to describe the insulin dynamics in a Type 1 Diabetic patient. The LTI model is the Ruan model, and a LSTM is trained to approximate the difference between the Ruan model and the real dynamics (from the UVA/Padova simulator). The objective is to build a model that is usable for real-time MPC, with fast inference time for optimization, but is still accurate. We also experiment with multi-step ahead prediction by concatenating multiple one-step ahead predictions in an autoregressive way. This is a project I would like to follow up on, for example experimenting with a transformer architecture for better multi-step ahead prediction.

Report ○ Presentation ○ Materials

Balancing and Trajectory Tracking for a Triple Inverted Pendulum on a Cart (Control Problems in Robotics - Underactuated Robots / April 2026) #

In this project we replicate results from two papers, exploring two different methods for trajectory tracking and balancing of a triple pendulum. The first method uses a BVP+LQR approach to generate and stabilize trajectories for the cart that result in desired motion of the pendulum links: with appropriate choice of boundary conditions and controller gains, both periodic motions and trajectories to switch between equilibrium points can be generated and tracked. The second method uses a Cascaded (recursive) version of External-Internal-Convertible form to track trajectories for the cart while keeping the links balanced in the upwards equilibrium.

Report ○ Presentation (pdf with videos) ○ Materials

RL-Augmented MPC for Humanoid Robot Locomotion (Autonomous & Mobile Robotics / March 2026) #

In this project we start from a humanoid robot, the HRP-4, which walks on flat ground using Model Predictive Control, and we explore the use of a residual policy, implemented as an agent trained with RL techniques, to be used alongside the MPC in order to achieve better rejection of perturbations, such as disturbance forces and terrain inclination. The simulation environment is implemented in DartPy, and the RL agent is trained using the PPO implementation in StableBaseline3.

Report ○ Presentation (pptx with videos)(odp with videos)(pdf)(videos) ○ Materials

Neural Lyapunov Control (Intelligent & Hybrid Control / January 2026) #

Using Neural Networks to learn controllers for dynamical systems and Lyapunov functions that prove the stability of the closed loop system.

Report ○ Presentation ○ Materials

Motion retargeting methods comparison (Medical Robotics / December 2025) #

In this project we implemented and compared different approaches to grasping tasks in a virtual reality environment, using the WeArt TouchDiver G1 haptic interface and the Meta Quest 3 VR headset. In particular, we replaced the TouchDiver animation in the WeArt SDK by modeling each finger in the hand as a planar 3R robot manipulator and implemented a grasping system based on physics interactions from scratch for the Meta SDK.

Report ○ Presentation ○ Materials

Model-based (sensorless) collision detection (Robotics 2 / September 2025) #

In this project we detect, isolate and identify collisions on a robot manipulator only using proprioceptive sensors. In particular, we leverage the concept of generalized momentum, which makes the task computationally efficient and suitable for real-time implementation.

Report (Addendum) ○ Presentation ○ Materials

BSc #

Control of a mobile robot with saturated inputs (Bachelor thesis / 18th October 2024) #

This is my bachelor thesis, written in Italian. I explored trajectory tracking using a mobile robot (unicycle/differential drive platform) via feedback linearization, with an additional correction to the control action provided by a MPC (Model Predictive Controller), to avoid joint saturation. I explore different cost functions and their impact on the actual trajectory traced by the robot with respect to the reference.

Thesis ○ Presentation ○ Materials

Two-wheel self-balancing robot (Automation / January 2024) #

This report is written in Italian, but I’ve written about the project in english in another post on this blog: a two-wheel self-balancing robot.

Report ○ Presentazione ○ Materials