Shared Voxel-Map Cooperative UAV Guidance with Multi-Agent Soft Actor-Critic
A study published on arXiv introduces a collaborative framework for guiding indoor UAVs, utilizing a shared voxel-map world model alongside a multi-agent Soft Actor-Critic (MASAC) controller. Various drones integrate 360-degree LiDAR data into a unified occupancy map, which is then transformed into a compact bird's-eye-view (BEV) format and given to each agent as an ego-aligned local segment. This design promotes consistent spatial fusion among multiple UAVs while allowing for decentralized continuous control. The policy incorporates BEV map characteristics, near-field obstacle data, and concise goal and peer-state information within a framework that centralizes training but decentralizes execution. In simulations, the developed controller achieved a 90.3% success rate in navigating corridors, surpassing A* planning, an artificial potential field controller, and previous guidance techniques. Additionally, the study tackles the issue of residual sim-to-real discrepancies.
Key facts
- The paper is on arXiv with ID 2607.25728.
- It proposes a cooperative indoor UAV guidance framework.
- The framework uses a shared voxel-map world model.
- It employs a multi-agent Soft Actor-Critic (MASAC) controller.
- Multiple drones fuse 360-degree LiDAR observations into a common world-frame occupancy map.
- The map is converted into a compact bird's-eye-view (BEV) representation.
- Each agent receives an ego-aligned local crop of the BEV.
- The design is integrate-in-world, act-in-ego.
- It enables consistent multi-UAV spatial fusion with decentralized control.
- The policy uses BEV map features, near-field obstacle observations, goal and peer-state information.
- Training is centralized, execution is decentralized.
- In simulation, success rate is 90.3% in corridor navigation.
- It outperforms A* planning, artificial potential field controller, and a prior guidance method.
- The paper addresses sim-to-real mismatch.
Entities
Institutions
- arXiv