Open-source, privacy-first multi-agent smart home system where autonomous software agents represent each occupant and device, negotiating resources peer-to-peer with no cloud dependency. The system learns routines implicitly through passive observation and adapts in real time to ecological data while keeping all sensitive data local.
Design, simulation and implementation of a mobile inverted pendulum using Model-Based Systems Engineering and the V-model. The cart autonomously balances a pendulum upright, rejects disturbances, recovers from initial tilts, and executes dynamic manoeuvres with two control strategies benchmarked end-to-end.
Modular, object-oriented grasp-planning pipeline that samples 6-DoF poses, executes grasps in PyBullet on multiple grippers and objects, and labels outcomes automatically. A random forest trained on pose features reached 82.9% validation accuracy and 60% success when tested on unseen simulation grasps.
Computational framework for optimising B-spline control points in acoustic levitation, incorporating smoothness, curvature and acoustic forces via the Gor'kov potential. Compared to linear interpolation, OptiSpline produces smoother trajectories with fewer oscillations, higher success rates and faster execution.
Derivation, implementation and calibration of a macroscopic delay differential equation model from the johnBuffer/AntSimulator agent-based model. Compares a full causal DDE approach against a hybrid ODE-DDE architecture to show foraging as stochastic search and homing as deterministic transport.
Six-wheeled ground robot built to autonomously navigate a multi-terrain obstacle course: line following, stair climbing, treadmill traversal, rough terrain, bridging surface gaps, and descending via a zipline — combining mechanical design, sensor integration and adaptive control.
Tabular Q-learning agent for a 4×4 four-in-a-row variant, trained through a curriculum of random-opponent play and self-play with symmetry reduction and ϵ-greedy decay. After 500,000 episodes it reaches ~0.98 win rate against random play; ablations and a 9×9 extension explore richer strategic depth.
Real-time 2D localisation of a mobile robot in a GPS-denied indoor arena using onboard IMU and Time-of-Flight sensors. An Extended Kalman Filter fuses inertial and range measurements to estimate position and yaw, evaluated against motion-capture ground truth on training trajectories.
Mathematical and mechanical analysis, prototyping and final assembly of a Theo Jansen–style Strandbeest. Includes MATLAB simulation of alternative walking trajectories, iterative prototyping, finite element analysis of components, and reflection on manufacturing lessons learned.
Cleaning, visualisation and next-day forecasting of NSW fuel price data (96k+ daily station–fuel records, 2016–2025). Explores cyclical price dynamics across fuel codes and builds per-station forecasting models from historical minimum prices.