DB Daniel Bryars academic file / msc-ai

MSc Artificial Intelligence / University of Leeds

MSc AI Work

MSc Artificial Intelligence, University of Leeds, completed 2026 with Distinction. Nine taught modules — programming, data science, algorithms, machine learning, knowledge representation, AI ethics, deep learning, text analytics and robotics — and a dissertation on robot learning.

These pages describe what I learned and show figures from my own assessments, notebooks and experiments. No Leeds teaching material or assignment briefs are reproduced.

Browse modules Robotics first
RViz screenshot from Daniel Bryars' robotics assessment work.
Robotics was where the MSc met the physical-world itch

Shape

The useful reset

The MSc gave me a structured way to revisit the maths, algorithms and research habits behind modern AI, after years of building production systems where reliability and judgement mattered every day.

Emphasis

Less magic, more mechanism.

The strongest through-line was learning to ask what the model is actually measuring, what the data represents, where the evaluation can lie, and what changes when the system leaves the notebook.

Range

From sandwiches to robots

The assessed work ranged from knowledge ontologies and AI ethics to orbital regression, image captioning, LLM summarisation proposals and reinforcement learning for simulated robots.

Result

What it changed

The programme sharpened the research side without replacing the engineering side: clean data, reproducible experiments, pragmatic tooling, careful evaluation and a preference for systems that can actually be made to work.

Gazebo world screenshot from the MSc robotics assessment.
Robotics world used for the navigation work
Learning curve from SARSA reinforcement learning in the robotics module.
SARSA learning curve
Mars orbit prediction plot from the machine learning module.
Mars position prediction
Looping Mars orbit plots from Gaussian process regression work.
Gaussian process regression, plotted until it made sense
Data science assessment output showing model metrics for fraud classification.
Fraud-classification metrics
Letter frequency chart from Programming for Data Science coursework.
Letter frequencies from the text pipeline
Protégé ontology screenshot from knowledge representation coursework.
Protégé ontology work
Encoder-decoder diagram from deep learning image captioning work.
Image-captioning model shape
Persona score table from data mining and text analytics coursework.
Persona variation experiment
Architecture diagram summarising practical MLX course projects.
MLX course project map

Dissertation

Robot learning, and whether it survives contact with hardware

I built a MuJoCo robot-arm simulation, connected it to a 3D headset through OpenVR to collect teleoperation data, then trained PyTorch policies on it — LeRobot, with ACT and ACT-ViT architectures — tracking experiments in W&B and keeping checkpoints on the Hugging Face Hub.

The part that mattered was testing those policies on the physical arm. Task success, the reality gap, resets, calibration, data quality and failure modes are where a policy that looks finished in simulation stops being finished.

The research manuscript is being reworked for ICRA 2027. The dissertation itself is not published here.

RViz view of the simulated robot and its sensor data during the MSc robotics work.
Simulation is the easy half

Taught modules

One page per module: what I studied, what I built, what I got wrong.

OCOM5100M / Python / data pipelines

Programming for Data Science

Project Gutenberg text processing, bulk data handling, letter statistics and a practical Python analysis pipeline.

OCOM5101M / classification / cost models

Data Science

Fraud prediction, model comparison, preprocessing, balanced error rates and business-facing cost analysis.

OCOM5102M / theory

Algorithms

Proof techniques, complexity, graph thinking and the useful discomfort of being precise.

OCOM5200M / regression / Gaussian processes

Machine Learning

Predicting Mars positions from JPL data using MLPs and Gaussian process regression.

OCOM5201M / ontology / reasoning

Knowledge Representation

A sandwich ontology, description logic, OWL reasoning, Lean and Prolog. Yes, sandwichness is surprisingly hard.

OCOM5202M / AI ethics

Ethics of AI

Facial recognition, surveillance, copyright, generative AI and the practical politics of powerful systems.

OCOM5203M / PyTorch / vision-language

Deep Learning

Image caption generation with ResNet encoders, sequence decoders, BLEU and cosine-similarity evaluation.

OCOM5205M / ROS / RL

Robotics

ROS, Gazebo, TurtleBot, laser scans, reward design and reinforcement learning for navigation.

MLX course / GitHub projects

MLX Course Projects

Six weeks of practical AI projects: vision transformers, captioning, audio, LoRA fine-tuning, embeddings and retrieval.