Information-Optimal Formation Geometry Design for Multimodal UAV Cooperative Perception

Kai Xiong, Xingyu Wu, Anna Duan, Supeng Leng, Jianhua HearXiv 2025 · DOI · arXiv

Abstract

Cooperative perception using swarms of Unmanned Aerial Vehicles (UAVs) is critical for applications like search and rescue and surveillance. However, the geometric arrangement of these swarms is often treated as a secondary byproduct of trajectory planning. This paper proposes a novel framework that treats 3D formation geometry as a primary design variable. By optimizing the allocation of heterogeneous sensors (LiDAR and cameras) and the spatial configuration of the swarm using information-theoretic principles, the proposed method achieves significant improvements in perception accuracy, field-of-view (FOV) coverage, and energy efficiency.

Introduction

In complex environments, a single UAV's perception is limited by its viewpoint and sensor type. Swarms overcome this by providing multi-angle, multimodal data. A key challenge lies in the fundamental differences between sensors: cameras provide high-resolution semantic information but lack depth precision, while LiDAR offers accurate 3D mapping at the cost of higher power and bandwidth.

Multimodal UAV swarm formation with LiDAR and Camera sensors.

Without strategic 3D positioning, swarms suffer from redundant data, blind spots, and communication interference. This research shifts the focus toward "formation-aware" design, where the swarm's shape is optimized to exploit sensor complementarity.

Methodology: A Three-Step Optimization Framework

The authors introduce a comprehensive framework to manage the complexity of multimodal cooperative perception, divided into three distinct phases.

The three-step optimization framework: UAV-Sensor Allocation, Formation Optimization, and Flight Control.

1. UAV-Sensor Allocation

The first step determines the optimal number of UAVs and the specific sensors they should carry. The authors utilize the Fisher Information Matrix (FIM) to quantify the information gain regarding a target's state. By maximizing the log-determinant of the FIM (D-optimality), the system identifies configurations that minimize estimation uncertainty. Because this objective function is submodular, a computationally efficient greedy algorithm can be employed to reach near-optimal solutions suitable for real-time deployment.

2. Formation Optimization

Once the sensor suite is defined, the swarm's spatial arrangement is optimized. The researchers introduce the concept of Equivalent Transformations—geometric flips that preserve the FIM determinant (and thus perception accuracy) while allowing the swarm to reconfigure for secondary objectives. These objectives include maximizing the collective Field-of-View (FOV) and improving the Signal-to-Interference-plus-Noise Ratio (SINR) for robust inter-UAV communication.

3. Flight Control

To transition the swarm into these optimal geometries, a distributed control scheme based on Lyapunov stability is used. A critical innovation here is the use of logarithmic potential fields instead of standard quadratic ones. Logarithmic fields bound the control authority, preventing the aggressive, high-energy maneuvers typical of traditional potential field methods, leading to smoother and more efficient flight paths.

Key Insights

  • Submodularity for Efficiency: Proving the submodularity of the FIM-based objective justifies the use of greedy search, making the complex combinatorial problem of sensor allocation tractable for onboard processors.
  • Sensor Complementarity: The framework explicitly places LiDAR-equipped drones in positions that best support camera-equipped drones, maximizing the strengths of each modality.
  • Energy-Aware Control: The shift to logarithmic potential fields directly addresses the physical constraints of UAV batteries, trading off aggressive convergence for long-term operational endurance.

Results and Evaluation

The framework was evaluated through high-fidelity numerical simulations of 3D target tracking. Compared to standard uniform formation benchmarks, the proposed approach yielded:

  • 25.0% improvement in FOV coverage, effectively eliminating blind spots.
  • 104.2% increase in SINR, ensuring reliable high-bandwidth data links.
  • 47.2% reduction in energy consumption, attributed to the optimized control laws.
  • Superior Accuracy: The multimodal (Camera + LiDAR) configuration achieved higher information gain (log-det(FIM) of 16.48) than single-modality swarms, even when using fewer total assets.

Conclusion

This work demonstrates that the 3D geometry of a UAV swarm is a powerful lever for enhancing cooperative perception. By integrating information theory with geometric optimization and energy-efficient control, the proposed framework provides a robust foundation for next-generation autonomous surveillance and search-and-rescue systems.

Summary updated

BibTeX
@misc{xiong2025information,
  author = {Kai Xiong and Xingyu Wu and Anna Duan and Supeng Leng and Jianhua He},
  title = {{Information-Optimal Formation Geometry Design for Multimodal UAV Cooperative Perception}},
  year = {2025},
  eprint = {2512.12923},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.12923},
}
Figure