David Cohen

M.Sc. Computer Science and Engineering

University of Bologna · Cesena

Weighted average 28.5/30, including 30 e lode in Machine Learning and Smart Vehicular Systems.

B.Sc. Computer Science and Engineering

University of Bologna · Cesena

Computer Engineering track. Thesis in natural language processing and decoding strategies for abstractive summarization.

Scala Robotics Simulator

A robotics simulator written in scala. In a team of three I designed the behaviour DSL and the domain validation model, built the illumination engine and the simulation GUI, and did the phototaxis learning task.

43M → 32states after redesigning the RL observation space

• ·Compose partial robot behaviours with Kleisli arrows, then close the chain with a fallback policy.
• ·Reduce an intractable Q-table to 32 states: the best checkpoint reached 90% success on ten held-out environments.
• ·Train a DQN on 10×10 arenas and evaluate on 5×5, 20×20 and 30×30 arenas: 86%, 95% and 100% success respectively on curated environments with different step budgets.

• python
• reinforcement-learning
• rl
• robotics

S-Parking

S-Parking: A modular Carla Simulator project for simulating parking scenarios with sensor fusion and state machine control. Two-person project. I authored the implementation.

7 / 24successful parks in a simulated evaluation; 17 safety stops, 0 collisions

• ·Give obstacle detection priority over manoeuvre completion through an immediate stop transition.
• ·Evaluate geometry, lateral offset and weather: 7 successful parks, 17 safety stops, zero collisions.
• ·Use an explicit rule-based controller. The evaluation is simulation-only; recovery from an interrupted manoeuvre remains a limitation.

ARGoS swarm robotics

Behaviour-based and swarm robotics in ARGoS/Lua: reactive control, subsumption, motor schemas and 50-robot aggregation. University of Bologna IRS course, 2024. Alone, for the course: the four controllers, the headless experiment scripts and the parameter study behind them.

81parameter combinations swept; the best set holds neighbours 27 cm apart

• ·Solve the same task three times over, with fixed priorities, with subsumption layers and with motor schemas summed as vectors, and compare what each architecture costs.
• ·Aggregate with a WALK/STOP automaton on ten-byte line-of-sight messages: stop with probability min(PsMax, S + α·N) in the number of stopped neighbours heard.
• ·Five hundred and three hundred headless runs with a Mann-Whitney test on the seeds; small parameter changes move the result a lot, and the report says so.

• lua
• multi-agent-systems
• potential-fields
• robotics

lecturize

Turn lecture recordings into notes offline: resumable faster-whisper CLI, markdown chapters/SRT/docx, local translation to 50 languages, no PyTorch. Mine, published on PyPI: design, implementation, packaging and releases. It replaces a single-file script I had been using for my own lectures.

7× / 20×faster than real time on my laptop, CPU then GPU, with the small model

• ·Drop PyTorch entirely and run CTranslate2: one install of about 300 MB, and the CUDA libraries only for whoever asks for them.
• ·Save every utterance to a SQLite checkpoint keyed by the file fingerprint and the settings, so an interrupted two-hour run continues from the last sentence.
• ·Translate with the Argos Translate model packages driven directly through CTranslate2, pivoting through English, instead of a second ML runtime.

• audio-transcription
• cli
• ctranslate2
• faster-whisper