Controlled and reproducible evaluation framework developed as part of the Master's thesis (University of Salerno), designed to assess zero-shot generalization of Vision-Language-Action (VLA) models on robotic manipulation tasks. The benchmark covers two generalization axes: Syntactic Generalization (verb substitution, syntactic restructuring, compositional spatial reference) and Task-Level Generalization (cross-object skill transfer, novel task composition), structured into levels of increasing difficulty on the LIBERO-Goal suite. Three state-of-the-art VLA models with distinct pre-training paradigms were evaluated: OpenVLA-OFT, TinyVLA, and InternVLA-M1, with targeted fine-tuning of InternVLA-M1 to mitigate catastrophic forgetting at the hardest compositional spatial reference level.

Autonomous indoor navigation system developed in ROS 2 on a TurtleBot 4 platform, as part of the Mobile Robots for Critical Missions course. The robot is tasked with reaching a target location via the shortest feasible path, complying with directional constraints imposed by colour-coded cones, and recovering from the kidnapped robot problem.

AI-assisted robotic system built on the Pepper platform for the Cognitive Robotics course, designed to enhance visitor experience in shopping mall environments. The system integrates a real-time video analytics module for face detection and tracking, a speech-to-text module based on Whisper (Azure OpenAI), and a conversational chatbot powered by OpenAI's Assistant API (GPT-4o) for multi-turn dialogue management. Quantitative evaluation of the chatbot yielded a Factual Accuracy of 0.92 on a dataset of 50 structured queries.

Computer vision system developed for the Artificial Vision Contest 2025 (University of Salerno), as part of a four-person team. The pipeline integrates YOLOv8-based person detection, multi-object tracking via BoT-SORT, individual attribute classification (gender, hat, and bag) using a trained MNAT classifier, and behaviour analysis based on crossings over configurable virtual lines.

Design and implementation of a hierarchical control system for a robotic arm trained to play ping-pong in a physics simulation, developed for the Machine Learning course (University of Salerno). The agent combines supervised pre-training of MLP-based inverse kinematics models (BaseMLP and ArmMLP) for long-range positioning with a DDPG (Deep Deterministic Policy Gradient) agent featuring a twin-critic actor-critic architecture (TD3-style) for fine-grained paddle control in the ball's proximity zone. Ball trajectory is predicted via quadratic ballistic equations; exploration is driven by Ornstein-Uhlenbeck noise with exponential sigma decay. The simulation environment runs in PyBullet with a decoupled client-server TCP architecture.

Systematic extension and improvement of an autonomous driving agent (BehaviorAgent) within the CARLA simulation environment, developed for the Autonomous Vehicle Driving course. The project addressed the key limitations of the baseline agent through dedicated modules for: lane change and safe overtaking with oncoming traffic detection, adaptive management of slow vehicles and cyclists, pedestrian and static obstacle detection and response, curve stability control, complex intersection handling, and stop sign recognition. The developed modules were validated on two predefined urban routes through quantitative comparison against the baseline agent.

Systematic analysis of the vulnerability of a face recognition system to adversarial attacks, carried out for the AI for Cybersecurity course. The baseline system (Inception ResNet V1 pre-trained on VGGFace2, 98.6% accuracy) was evaluated against five attack families in both error-generic and error-specific modes: FGSM, BIM, PGD, DeepFool, and Carlini & Wagner L∞. Attack transferability to a second model (ResNet-50) was assessed in a grey-box scenario, and two defence strategies were designed and evaluated: preprocessing-based defences (JPEG compression, feature squeezing, spatial smoothing) and explicit adversarial sample detection using a fine-tuned ResNet-50 detector.

Design and implementation of a firmware for a dual-leaf electronic gate, developed for the Embedded Systems course using a model-based development approach. The control logic is modelled as a hierarchical state machine in MATLAB Stateflow, handling gate opening and closing, obstacle detection via proximity sensors, hardware button debouncing, configurable movement and auto-close timers, LED feedback, and error state management. C code was automatically generated via Embedded Coder and validated against 19 test scenarios using MATLAB Test Harness before deployment on the STM32 NUCLEO-G474RE board.