Safe Cooperative Decision-Making for Multi-UAV Pursuit-Evasion Games via Opponent Intent Inference.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Cooperative multi-unmanned aerial vehicle (UAV) pursuit-evasion faces challenges from occlusions, sensor noise, and non-stationary evader tactics.
• These factors lead to intermittent observability, destabilize predictions, and compromise safety-constrained cooperation.

Purpose of the Study:

• To develop a safe decision-making framework for cooperative multi-UAV pursuit-evasion that enhances interpretability and uncertainty awareness.
• To address challenges of intermittent observability, varying observation window lengths, and non-stationary evader tactics.

Main Methods:

• Proposed a generative intent-subgoal model for evader behavior mode and subgoal inference from short observations.
• Developed a length-agnostic trajectory predictor using multi-window knowledge distillation for uncertainty-aware predictions.
• Implemented a belief-risk-gated hierarchical multi-agent policy with a safety projection layer for adaptive strategy switching.

Main Results:

• Demonstrated more stable cooperative capture, safer maneuvering, and lower decision variance in obstacle-rich environments.
• Achieved a 5-7% improvement in normalized expected return and reduced pursuer losses by 22-25% across different observation settings.
• Maintained end-to-end decision latency within a 50 ms control cycle, showing real-time feasibility.

Conclusions:

• The proposed framework offers robust and safe cooperative pursuit-evasion capabilities in challenging, dynamic environments.
• The interpretable, uncertainty-aware approach enables adaptive strategy and controllable trade-offs between efficiency and safety.
• The method shows strong real-time performance and superiority over existing baselines.