Collaborative Target Search With a Visual Drone Swarm: An Adaptive Curriculum Embedded Multistage Reinforcement...

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Collaborative target search (CTS) with visual drone swarms is crucial for disaster rescue and logistics.
• Challenges include 3D sparse reward exploration, limited visual perception, and complex collaborative behaviors.
• Existing methods struggle with data efficiency and adaptability in dynamic environments.

Purpose of the Study:

• To develop a data-efficient deep reinforcement learning (DRL) approach for CTS in visual drone swarms.
• To address challenges of sparse rewards, limited perception, and multi-agent collaboration.
• To enable effective autonomous CTS operations in real-world scenarios.

Main Methods:

• Proposed adaptive curriculum embedded multistage learning (ACEMSL) for visual drone swarms.
• Decomposed CTS into subtasks: obstacle avoidance, target search, and inter-agent collaboration.
• Implemented an adaptive embedded curriculum (AEC) to adjust task difficulty based on success rate.

Main Results:

• ACEMSL demonstrated data-efficient training and effective individual-team reward allocation.
• The approach was successfully deployed on a real drone swarm for CTS without fine-tuning.
• Extensive simulations and real-world tests validated the method's effectiveness and generalizability.

Conclusions:

• ACEMSL provides an effective solution for data-efficient collaborative target search in visual drone swarms.
• The method addresses key challenges in sparse reward exploration and multi-agent coordination.
• Validated effectiveness in both simulated and real-world flight tests, paving the way for practical applications.