In my 2nd year at Télécom Paris, I took part in an 4-months AI challenge called "Swarm rescue", with my teammate Erwan Fagnou. The challenge was organized by the DGA (a part of the French army).
Here is the original description of the challenge:
"With this project, you will try to save human lives, in simulation... Teach a swarm of drones how to behave to save a maximum of injured people in a minimum of time!
The objective of the mission is to map an unknown and potentially hostile area, detect targets (injured or healthy people) and guide them out of the area. A typical use case is the investigation of the basement of a collapsed building in the dark, in order to locate trapped people and rescue them.
Each team will have a fleet of 10 drones. Each drone will be equipped with communication and sensor capabilities.
Your job is to make these drones fully autonomous by programming them in Python.
The drones will have to manage the limited range of communication means, collaborate with each other to acquire information and relay it to an operator outside the area, be able to handle sensor and communication failures and unforeseen events such as the loss of the drones in order to conduct this mission autonomously.
The work on the challenge will be done exclusively in this simulation environment, with maps of increasing complexity. The final evaluation will be done on an unknown map made available to each team at the time of the demonstration. Each proposal will be tested on the same computer and a score related to the performance will be calculated.
I made a video presentation explaining our solution. The video is in french, but with "Auto-translate" english subtitles available. I found that the auto-subtitles are pretty good, please e-mail me if you'd like that I add english subtitles manually.
The video lasts 25mn, and is a captation of the same technical presentation we made in front of the jury. We got a grade of 20/20 for this presentation (see Results below).
My teammate Erwan and I won the Swarm Rescue 2021 challenge. Here are the detailed final results. We got 72.8/100 for the technical score (reward optimized by our RL agent) across 10 runs. The second and third solutions (by students from ENSTA and Supaero) got 47.9% and 26.9%.
We got 20/20 for the presentation score, including 5/5 on "solution credibility", 5/5 on "presentation quality", 5/5 on "question answering".
The prize was 8,000€.
Our code is not available publicly. We may give you read rights case-by-case on demand.
Here are the slides used during the technical presentation (in french):