Multi Agent Framework for Collective Intelligence Research
Alexandru Dochian
TL;DR
This work introduces a scalable, decentralized multi-agent framework for collective intelligence research that enables networked information exchange among autonomous units in a 2D environment with local sensing. It combines a modular architecture (AppProcess, Controllers, Environments, LogHandlers) with field modulation to construct local perception maps guiding collision avoidance and gradient-based behavior, validated through Hello World and drone experiments in the VU Amsterdam lab. The approach demonstrates sim-to-real consistency across simulation and physical Crazyflie drones and emphasizes configuration-driven replication via JSON, setting the stage for on-board processing and cloud-scale extensions. Overall, it provides a flexible platform for developing and testing distributed decision-making in swarm-like systems with practical pathways toward addition of deep reinforcement learning controllers and broader hardware integration.
Abstract
This paper presents a scalable decentralized multi agent framework that facilitates the exchange of information between computing units through computer networks. The architectural boundaries imposed by the tool make it suitable for collective intelligence research experiments ranging from agents that exchange hello world messages to virtual drone agents exchanging positions and eventually agents exchanging information via radio with real Crazyflie drones in VU Amsterdam laboratory. The field modulation theory is implemented to construct synthetic local perception maps for agents, which are constructed based on neighbouring agents positions and neighbouring points of interest dictated by the environment. By constraining the experimental setup to a 2D environment with discrete actions, constant velocity and parameters tailored to VU Amsterdam laboratory, UAV Crazyflie drones running hill climbing controller followed collision-free trajectories and bridged sim-to-real gap.