Communication-Free Collective Navigation for a Swarm of UAVs via LiDAR-Based Deep Reinforcement Learning

TL;DR

The paper tackles the challenge of coordinating a UAV swarm in environments where communication and global localization are unavailable. It proposes a fully LiDAR-driven, DRL-based controller that enables implicit leader-following: followers learn cohesive flocking and obstacle avoidance policies using only onboard perception, with one leader providing goal information. The approach combines a LiDAR perception stack (LIO, DBSCAN clustering, EKF tracking, and cluster validation) with a PPO-trained policy that processes ego-state, a fixed number of neighboring observations, and a 2-channel occupancy grid to output 3-D velocity commands. Demonstrations in extensive simulations and real-world experiments with five UAVs show robust performance, superior baselines comparison, and successful sim-to-real transfer in indoor and outdoor settings. The work advances practical, scalable, and resilient collective navigation for communication-denied swarms, paving the way for larger-scale deployments and adaptive role strategies.

Abstract

This paper presents a deep reinforcement learning (DRL) based controller for collective navigation of unmanned aerial vehicle (UAV) swarms in communication-denied environments, enabling robust operation in complex, obstacle-rich environments. Inspired by biological swarms where informed individuals guide groups without explicit communication, we employ an implicit leader-follower framework. In this paradigm, only the leader possesses goal information, while follower UAVs learn robust policies using only onboard LiDAR sensing, without requiring any inter-agent communication or leader identification. Our system utilizes LiDAR point clustering and an extended Kalman filter for stable neighbor tracking, providing reliable perception independent of external positioning systems. The core of our approach is a DRL controller, trained in GPU-accelerated Nvidia Isaac Sim, that enables followers to learn complex emergent behaviors - balancing flocking and obstacle avoidance - using only local perception. This allows the swarm to implicitly follow the leader while robustly addressing perceptual challenges such as occlusion and limited field-of-view. The robustness and sim-to-real transfer of our approach are confirmed through extensive simulations and challenging real-world experiments with a swarm of five UAVs, which successfully demonstrated collective navigation across diverse indoor and outdoor environments without any communication or external localization.

Figures (11)

• Figure 1: An example of real-world validation of the proposed LiDAR-based collective navigation. (a) A swarm of five UAVs on standby for takeoff. (b) Trajectories of the swarm reconstructed from onboard LiDAR. (c) Sequential snapshots of the swarm avoiding obstacles to reach a goal without external localization or communication. More details can be found in the attached video at https://youtu.be/U4i3Spisugg.
• Figure 2: Scenario of the communication-free collective navigation problem. A leader visits a known waypoint sequence, while followers, unaware of the goal or leader's identity, use only local perception for flocking and obstacle avoidance to indirectly follow the leader.
• Figure 3: Overview of the onboard perception system. (a) The perception pipeline. (b) The object tracker filters raw LiDAR points, clusters them using DBSCAN, tracks them with an EKF, and validates them based on temporal consistency.
• Figure 4: The proposed DRL architecture. An encoder processes observations into a latent vector, and actor and critic heads use it to determine the policy and estimate value.
• Figure 5: Illustration of reward function components for learning communication-free collective navigation. (a) Flocking: cohesion encourages staying close to neighbors, while separation maintains safe distances. (b) Obstacle avoidance: proximity penalizes closeness to obstacles, and direction penalizes movement toward obstacles. (c) Stable flight: altitude maintenance encourages following the leader's altitude, and attitude stability promotes upright orientation. (d) Neighbor perception: visibility rewards keeping neighbors within FOV, recovery triggers descent when all neighbors are lost.