Joy Baroudi

Designing, implementing and evaluating an end to end vision language action pipeline that lets a Unitree G1 humanoid carry out household assistance tasks from natural language instructions. Demonstrations are collected by teleoperation with VR and RGB-D cameras, converted into a common VLA training format and used to fine tune open models such as Pi0 and Pi0.5. The work is aimed at safe, reliable execution on the real robot with a person nearby, measured on success rate, robustness, generalisation and how often a human has to step in.

A detection model that reads a surgical tray and names what is on it, so an assisting system can tell at a glance what is present and what is missing. Trained to separate instruments that look almost identical: curved and straight Mayo scissors, a scalpel and a dissection clamp, on reflective steel under theatre lighting.

A compact retrofit that opens a normal interior door for people with limited mobility. High torque actuators and a servo latch release driven from an Arduino, with limit sensing and a force cut-off so it stops if something is in the way. Built alongside a small market-fit study.

A tracked robot that searches a space, finds a flame and puts it out with no operator in the loop. IR sensors handle obstacle avoidance and environment sensing, and the decision logic arbitrates between searching, approaching and extinguishing depending on what the sensors currently agree on.