Fast Swarming of UAVs in GNSS-denied Feature-poor Environments without Explicit Communication

TL;DR

This work tackles fast multi-UAV swarming in GNSS-denied, feature-poor environments using onboard sensing only. It introduces a decentralized, state-feedback flocking controller tied to an enhanced Multi-Robot State Estimator (MRSE) that fuses with VIO, a neighborhood model for robust mutual perception via UVDAR, and an adaptive fusion scheme with a communication-less variant that estimates neighboring states to reduce dependence on unreliable networks. Key contributions include the distributed flocking controller, enhanced MRSE, neighborhood-based perception, and velocity estimation without explicit inter-agent communication, all validated in real-world experiments reaching the hardware limits of group speed. The approach delivers improved speed, resilience to GNSS jamming or spoofing, and SPoF resistance, enabling reliable outdoor-indoor transitions and scalable large-swarm deployment in challenging environments.

Abstract

A decentralized swarm approach for the fast cooperative flight of Unmanned Aerial Vehicles (UAVs) in feature-poor environments without any external localization and communication is introduced in this paper. A novel model of a UAV neighborhood is proposed to achieve robust onboard mutual perception and flocking state feedback control, which is designed to decrease the inter-agent oscillations common in standard reactive swarm models employed in fast collective motion. The novel swarming methodology is supplemented with an enhanced Multi-Robot State Estimation (MRSE) strategy to increase the reliability of the purely onboard localization, which may be unreliable in real environments. Although MRSE and the neighborhood model may rely on information exchange between agents, we introduce a communication-less version of the swarming framework based on estimating communicated states to decrease dependence on the often unreliable communication networks of large swarms. The proposed solution has been verified by a set of complex real-world experiments to demonstrate its overall capability in different conditions, including a UAV interception-motivated task with a group velocity reaching the physical limits of the individual hardware platforms.

Figures (9)

• Figure 1: A swarm of six UAVs (yellow) tracking an intruder drone (red) using the proposed approach. The group velocity of 5ms was reached, while the swarm stayed coherent without reliance on GNSS and communication. Circles show the positions of UAVs at the beginning of the experiment.
• Figure 2: a) The system architecture for a single robot. High-level Flocking controller determines the desired velocity of UAV based on the neighborhood state (modeled in Data collector) and relative position of the goal (e.g., tUAV, in the interception-motivated task). The red arrows show the data flow in the process of the proposed neighbors' velocity estimation approach. In the Enhanced MRSE block, VIO is fused with states of the enhanced cooperative estimator. b) Visualization of the desired position in the formation, group heading, geometrical formation rules, distance to target limits, and frames of reference.
• Figure 3: Example of an oUAV's neighborhood estimation. oUAV's neighborhood may not be fully observable by fUAV.
• Figure 4: Top view of swarm approaching a static goal with group velocity 5ms. The red circle is a motion reference.
• Figure 5: Hardware setup of UAVs used in experiments.