AttentionSwarm: Reinforcement Learning with Attention Control Barier Function for Crazyflie Drones in Dynamic Environments

TL;DR

AttentionSwarm addresses safe, scalable multi agent drone control in dynamic environments by fusing a centralized MAPPO policy with an attention based Control Barrier Function. The Attention CBF network yields adaptive safety penalties that constrain the learned policy while preserving performance, enabling real time collision avoidance. Through extensive simulation and indoor real world experiments with Crazyflie drones, the approach achieves 95-100% collision free navigation and outperforms baselines such as MAPPO, A2C, and MPC. The work provides a practical framework for safe high speed multi robotic operations in applications like logistics, inspection, and racing.

Abstract

We introduce AttentionSwarm, a novel benchmark designed to evaluate safe and efficient swarm control in a dynamic drone racing scenario. Central to our approach is the Attention Model-Based Control Barrier Function (CBF) framework, which integrates attention mechanisms with safety-critical control theory to enable real-time collision avoidance and trajectory optimization. This framework dynamically prioritizes critical obstacles and agents in the swarm's vicinity using attention weights, while CBFs formally guarantee safety by enforcing collision-free constraints. The AttentionSwarm algorithm was developed and evaluated using a swarm of Crazyflie 2.1 micro quadrotors, which were tested indoors with the Vicon motion capture system to ensure precise localization and control. Experimental results show that our system achieves a 95-100% collision-free navigation rate in a dynamic multi-agent drone racing environment, underscoring its effectiveness and robustness in real-world scenarios. This work offers a promising foundation for safe, high-speed multi-robot applications in logistics, inspection, and racing.

Figures (5)

• Figure 1: System overview of AttentionSwarm, a reinforcement learning-based system enabling drone swarms to safely navigate in dynamic environments with moving obstacles and changing conditions. The system is evaluated across a dynamic drone racing scenario.
• Figure 2: Gym-PyBullet simulation setup with a moving obstacle.
• Figure 3: Recorded trajectories of two drones during a real-flight racing scenario. The drones navigate a course containing a green square gate and a yellow cylindrical obstacle, demonstrating agile flight and obstacle avoidance in a physical environment.
• Figure 4: Evaluating swarm agility and path planning in complex simulation environments. Colored trajectories illustrate drones navigating two challenges: (a) and (b) flying straight through cluttered obstacles, and (c) and (d) coordinating in a complex racing track.
• Figure 5: Convergence of the learning algorithm. The policy demonstrates stable learning, with rewards increasing and episode times shortening after 15,000 epochs, while the explained variance remains stable near 1.0, indicating successful training.