Neural Surrogate-Enhanced Metaheuristic Optimization for Distributed Quadrotor Swarm Control.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Real-time cooperative control of quadrotor swarms in cluttered environments presents challenges in balancing formation, obstacle avoidance, safety, and computational cost.
• Existing methods often struggle to efficiently manage these competing objectives in dynamic, complex scenarios.

Purpose of the Study:

• To propose a multilayer perceptron (MLP) surrogate for high-level objective-weight selection in a modified multi-objective pigeon-inspired optimization (MPIO) controller.
• To enhance the performance of quadrotor swarm control by learning state-to-weight mappings for direct objective-weight vector prediction.

Main Methods:

• Developed an MLP surrogate trained using teacher-generated weight labels from randomized scenes, DAgger-based state aggregation, and risk-weighted supervision.
• Integrated the MLP surrogate into a modified MPIO distributed controller, retaining core flocking, obstacle-avoidance, and command generation rules.
• Evaluated the controller on a fixed closed-loop benchmark and a qualitative AirSim case study.

Main Results:

• Significantly increased the true collision-free rate from 48.00% to 86.89% and the safe success rate from 38.67% to 74.22% compared to the baseline modified MPIO.
• Drastically reduced mean per-step decision latency for the swarm from 8494.70 ms to 0.92 ms.
• Demonstrated that the MLP surrogate provides a superior safety-latency tradeoff, with medium-sized networks performing best.

Conclusions:

• The proposed MLP surrogate effectively optimizes quadrotor swarm control for cluttered environments, prioritizing safety and runtime efficiency.
• The approach shows high-fidelity simulation transferability, indicating potential for real-world deployment in complex navigation tasks.
• The MLP surrogate offers a significant advancement in real-time cooperative control for multi-UAV systems.