Deep reinforcement learning for time-critical wilderness search and rescue using drones

Area of Science:

• Robotics
• Artificial Intelligence
• Search and Rescue Technology

Background:

• Traditional wilderness search and rescue (SAR) methods are often slow and cover limited areas.
• Drones offer a more efficient and adaptable solution for SAR operations.
• Optimizing drone flight paths is critical for maximizing search effectiveness.

Purpose of the Study:

• To develop a novel algorithm for creating efficient drone search paths in wilderness environments using deep reinforcement learning.
• To leverage a priori data, including probability distribution maps of the search area and missing persons, to guide the learning process.
• To enhance the speed and probability of locating missing individuals in SAR operations.

Main Methods:

• A deep reinforcement learning algorithm was employed to train drones for optimal search path planning.
• The algorithm utilizes probability distribution maps to inform the drone's policy for efficient searching.
• A continuous action space, enabled by cubature, was incorporated for more sophisticated flight patterns.

Main Results:

• The proposed algorithm demonstrated a significant reduction in search times compared to traditional methods.
• Experimental results showed improvements exceeding traditional coverage and search planning algorithms.
• The use of a continuous action space allowed for more nuanced and effective flight maneuvers.

Conclusions:

• Deep reinforcement learning provides an effective solution for optimizing drone search paths in wilderness SAR.
• The algorithm's ability to leverage prior data and employ nuanced flight patterns leads to substantial improvements in search efficiency.
• This technology has the potential to significantly enhance the speed and success rates of real-world search and rescue missions.