Matteo Bando

We present YOLODOOM, a real-time vision-based system for enemy detection and automated aiming in Doom, demonstrating that raw screen output alone — operating on an unmodified game executable via OS-level screen capture and input emulation, requiring no simulator, no custom API, and no engine integration — is enough to drive real-time autonomous aiming in a classic FPS environment. This positions our system closer to a human-playable setting than prior ViZDoom-based approaches, which rely on privileged environment access. We evaluate two detection approaches: a multi-scale template matching baseline and a YOLOv8n model trained on CocoDoom. The YOLO detector reaches about 106 FPS with 6 ms inference time, outperforming template matching by 4.6× in throughput and 6.3× in inference speed, and remaining stable across all tested confidence thresholds. We further analyze the aiming behaviour of the controller against a real human player: despite being fully deterministic, it is qualitatively comparable to human performance, with overshoot rates of 48.5% (bot) vs 53.6% (human). The underlying architecture is game-agnostic and extensible to other titles through domain-specific fine-tuning.

Deep learning models for medical image classification achieve high accuracy but rely on large, costly labeled datasets. Active Learning (AL) mitigates this by iteratively selecting informative samples under a fixed labeling budget. However, batch-mode AL often struggles with redundancy when relying solely on scalar uncertainty scores. In this work, we formulate batch-mode AL for skin lesion classification as a Multi-Objective Optimization (MOO) problem, explicitly balancing informativeness, coverage, and diversity. We employ bio-inspired optimizers, NSGA-II with set-based encoding and MOPSO with continuous scoring vectors, to search for Pareto-optimal batches. We evaluated our MOO approaches on the HAM10000 dataset, where they were able to outperform established baselines in sample efficiency, with MOPSO attaining the highest Area Under the Learning Curve (AULC). Furthermore, Pareto-front analyses provide interpretable insights into the trade-offs between the competing objectives. This work demonstrates that evolutionary multi-objective algorithms offer a superior, transparent approach to resolving conflicting acquisition criteria in deep Active Learning.

We analyze whether the localization abilities that emerge in dino.txt correspond to genuine compositional grounding. The model shows strong object–attribute binding and contextual robustness, but fails systematically under relational and priority-based role inversions — revealing a dissociation between localizing an object and understanding the composition that describes it.

We investigate how to improve CLIP on base classes from a handful of labelled samples while preserving its zero-shot generalization to novel ones. Few-shot adaptation is combined with zero-shot CLIP predictions through a prototype-guided balancing mechanism, letting the model decide which signal to trust per sample. Zero-shot performance is preserved, base-class accuracy improves by +15–20%, and the harmonic mean rises from 74.6% to 82.7%.