Abstract
Designing autonomous aerial robot team systems remains a grand challenge in robotics. Existing works in this field can be categorized as centralized and decentralized. Centralized methods suffer from scale dilemmas, while decentralized ones often lead to poor planning quality. In this paper, we propose an enhanced decentralized autonomous aerial robot team system with group planning. According to the spatial distribution of agents, the system dynamically divides the team into several groups and isolated agents. For conflicts within each group, we propose a novel coordination mechanism named group planning. The group planning consists of efficient multi-agent pathfinding (MAPF) and trajectory joint optimization, which can significantly improve planning quality and success rate. We demonstrate through simulations and real-world experiments that our method not only has applicability for a large-scale team but also has top-level planning quality.
Introduction
Autonomous aerial swarms face a fundamental trade-off between optimality and scalability. Centralized methods provide high-quality, collision-free coordination but scale poorly as the number of agents increases. Conversely, decentralized systems offer excellent scalability but often struggle in dense environments, leading to suboptimal trajectories or navigation failures in tight spaces. This paper introduces a hybrid "group planning" framework that bridges this gap by dynamically grouping drones in close proximity for joint trajectory optimization while maintaining decentralized operation for isolated agents.
Methodology
The proposed system utilizes an "on-demand centralization" approach. Drones operate independently until they enter a defined safety threshold, at which point they form a temporary group to resolve potential conflicts.
The pipeline consists of three primary stages:
- Dynamic Grouping: Agents monitor neighbors and form groups when spatial density increases.
- Efficient Multi-Agent Pathfinding (EMAPF): Within a group, an improved ECBS algorithm with a 3D gridmap tie-breaker finds initial collision-free paths in milliseconds.
- Group Trajectory Joint Optimization: Using the MINCO trajectory class, the group performs joint optimization. A novel gradient-based coordination mechanism accounts for reciprocal avoidance, ensuring trajectories converge to a joint optimum.
Experimental Results
The system was evaluated against state-of-the-art frameworks including EGO-Swarm and DSTO across simulated and real-world scenarios.
Narrow Gate Navigation
In a challenging test requiring multiple drones to pass through a narrow aperture, the proposed method demonstrated superior reliability. While competing decentralized methods like EGO-Swarm and DSTO experienced collisions due to a lack of coordinated planning, the proposed group planning approach achieved a 100% success rate.
Large-Scale and Real-World Performance
- Efficiency: In an 8-drone circle swap, the proposed method achieved a flight time of 20.75s, significantly faster than EGO-Swarm's 28.48s.
- Stability: The system required only 98 replans compared to over 19,000 for MADER, indicating much higher trajectory confidence.
- Scalability: Simulations with up to 50 drones confirmed the system's ability to handle complex local conflicts through 35 dynamic group planning events.
Conclusion
By combining the strengths of centralized coordination and decentralized scalability, the group planning framework significantly enhances the performance of aerial swarms in cluttered environments. Future work will focus on integrating kino-dynamic constraints more deeply into the initial pathfinding phase and improving robustness against communication delays in the broadcast network.
Summary updated