FlightMARL: A Multi-Agent Reinforcement Learning Framework for Vision-Based Control of Autonomous Quadrotors

Abstract

In this paper, we present FlightMARL, an extension of the Flightmare simulator that implements a modular multi-agent reinforcement learning engine capable of supporting an array of models and algorithms. We explore representation learning of various different models in this environment. We implement recurrent agents as both discrete and continuous networks. We evaluate the learned representations of the models in the multi-agent environment. We demonstrate that agents constructed from continuous-time neural networks achieve interpretable causality in their representations.