Multi-UAV task allocation for solar-plant inspection
A fleet of drones splits the reconnaissance of a photovoltaic plant between them. Agents were trained with maskable and recurrent PPO and evaluated in AirSim.
- Python
- PPO
- AirSim
- Multi-agent RL
AI and software engineer in Utrecht. I train reinforcement-learning agents to plan flight paths, and write the software that gets them onto real drones.
Open to internships, thesis collaborations and engineering roles
Running a business?AI automations from €750
Click the field to move the goalTap the field to move the goal
A grid world, solved live in your browser with value iteration. The orange ring is the goal and is worth 1. Every step away from it is discounted by γ = 0.97, so each dot’s size is how valuable that cell is.
Each Bellman sweep pushes value one cell further out. When the wave reaches the drone, it follows the greedy policy: always fly to the neighbour worth the most. The hatched blocks and this page’s text are walls. Click anywhere in the field to move the goal.Tap anywhere in the field to move the goal.
01 / Work
Research, team and personal projects from the last two years, across the four areas I spend most of my time on. Pick one to filter.
Custom Gymnasium environments, recurrent and maskable PPO, and reward design for 3D path planning and multi-agent task allocation.
Fiducial tag detection with OpenCV, built to run on board a drone’s companion computer in real time.
PX4 and ArduPilot over MAVLink, ROS 2 and Gazebo, containerised so a full flight stack starts with one command.
Full-stack work in TypeScript, Vue, PHP and Python, and MCP servers that let language models use real tools.
A fleet of drones splits the reconnaissance of a photovoltaic plant between them. Agents were trained with maskable and recurrent PPO and evaluated in AirSim.
OpenCV tag detection for autonomous missions, plus a Docker setup that brings up PX4 SITL, Gazebo and MAVLink together, so every team member tests against the same simulator.
Model Context Protocol servers that give language models structured, permissioned access to company data and actions, alongside full-stack product work.
BSc thesis, Utrecht University
A custom training environment in which a quadcopter learns to reach goals through cluttered 3D space. Compared recurrent and maskable PPO, then validated the policies in Gazebo with ArduPilot over MAVLink.
Personal project
Several airframe iterations designed in Fusion 360, 3D-printed and assembled at home. The current 10-inch platform flies stably and carries a companion computer for onboard autonomy.
02 / About
I’m an MSc student in Artificial Intelligence & Engineering Systems at Eindhoven University of Technology, and a computer-vision engineer at Drone Team Aero. My work sits where learning meets hardware: agents that plan in three dimensions, perception that runs on a companion computer, and the simulation stacks that let you test both before anything leaves the ground.
Before that I spent two and a half years shipping web products and automation in Python, PHP and TypeScript. I still care about that side of the craft: clear interfaces, dependable pipelines, and code the next person can read.
03 / CV
Mar 2026 – Present
Drone Team Aero, Eindhoven
Nov 2025 – Apr 2026
CleverMobi, Eindhoven
Jun 2023 – Dec 2025
VMNmedia, Zeist
Feb 2015 – May 2023
Hospitality and sport
2025 – 2027
Eindhoven University of TechnologyGPA 8.2
Coursework: ROS 2, Systems & control, Physics, Deep learning
2022 – 2025
Utrecht UniversityGPA 7.4
Coursework: Machine learning, Calculus, Programming, Logic, Research, Writing, Presenting
04 / Contact
Open to internships, thesis collaborations and engineering roles in machine learning, robotics and UAVs. Email is the fastest way to reach me.
nemoterbogt@gmail.com