Decentralised Coordination for a Quadcopter Swarm
Netherlands
A runtime for collective-intelligence experiments, where each agent decides from a local perception map and nothing holds a global view. Tuned in simulation, then flown on Crazyflie quadcopters.
The situation
A swarm has no coordinator. Each unit sees a fraction of the situation and has to choose an action that does not collide with the choices its neighbours are making at the same moment, from information that is already slightly stale. The research question was whether behaviour tuned in simulation survives contact with real hardware.
What we built
The runtime
Agents, controllers, environments and loggers are independent processes that exchange state over a Redis channel. An experiment is a JSON configuration rather than a code change, so a run is reproducible and its parts are reusable across experiments.
Local perception
Each agent builds an 84x84 map of what it can sense, itself at the centre, from neighbour positions and the points of interest the environment publishes. The map is pooled to a 3x3 grid and the highest-scoring cell becomes one of nine discrete actions. No agent ever holds the global picture.
Simulation to hardware
Parameters were tuned against virtual agents, then the same configuration drove five Crazyflie quadcopters over radio in a laboratory. Swapping the agent implementation was the only change, because the controller never knew which kind it was steering.
What mattered
Keeping the controller ignorant of whether it drove a simulated or a physical agent is what made the transfer cheap. The interface was the experiment: had the controller known, every parameter tuned in simulation would have needed retuning against hardware.
Outcome
Five quadcopters held collision-free trajectories under rotating points of interest, following in the laboratory the behaviour seen in simulation. The framework and the paper are public.