MARL for Swarm Communication
A multi-agent reinforcement learning framework for learning abstract communications with DOTS swarm robots.
Exploratory work during the taught component of the FARSCOPE CDT (Sept 2024), building a simulation framework for running efficient MARL experiments for the Bristol swarm group’s DOTS robots.
- Developed a novel multi-agent scenario in which agents must learn to correctly mix paint colours based on communication accuracy and interpretation.
- Evaluated a broad range of MARL algorithms and deep learning network architectures across the scenario.
- Built an efficient simulation framework tailored to the constraints of real swarm hardware.