Distributed vision-only cooperative flight of multiple quadrotors in unknown cramped environments

Multi-quadrotor systems hold substantial promise for various applications, yet navigating autonomously in complex and unknown environments like forests and alleyways remains a formidable challenge. Contemporary multi-UAV systems predominantly rely on wireless communication for navigation in environments with known or no obstacles. Further extensive research is imperative to enhance the perception, cooperation, and control of multi-UAV systems. This paper delineates the development of a distributed vision-only multi-UAV system, wherein the agents exclusively depend on onboard sensing and computation for navigation and cooperative flight in unknown, constrained environments. The onboard sensors comprise a frontdepth camera and an omnidirectional gray monocular camera. The end-to-end sensorimotor controller computes the control commands, effectuating a direct mapping from perception to action through a sophisticated neural network design. This design is trained in simulation via imitation learning, namely, by emulating an expert endowed with access to privileged information. Such a setup empowers agents to accomplish cooperative formation flight, obstacle avoidance, target tracking, and other maneuvers in constrained environments, independent of wireless communication. Our findings underscore that distributed visiononly policies can rival traditional and other learning-based autonomous flight methodologies in challenging environments.