AERIS: Offline Policy Improvement for Multi-UAV Integrated Sensing and Communication
This paper proposes AERIS, an offline policy improvement framework featuring the STAR-CRDT algorithm, which enables safe and efficient multi-UAV integrated sensing and communication control by learning from historical flight logs to significantly boost performance metrics while eliminating risky online trial-and-error exploration.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the near future, the wireless networks that connect our devices will do more than just transmit data; they will also act as eyes and ears for the physical world. This emerging concept, known as integrated sensing and communication, envisions a single system that can talk to your phone while simultaneously mapping its surroundings, detecting moving objects, and monitoring the environment. When unmanned aerial vehicles, or drones, are added to this mix, the possibilities expand significantly. These flying machines can hover over areas where ground towers cannot reach, creating clear lines of sight to users and targets. However, controlling a fleet of drones to perform both tasks at once is a formidable challenge. The drones must constantly adjust their positions and the direction of their signals to maintain high-speed internet connections and accurate sensing, all while avoiding collisions with each other and ensuring they do not lose track of the objects they are monitoring. Because the environment is unpredictable and the targets move randomly, finding the perfect flight path for every drone in real-time is incredibly difficult and computationally expensive.
Researchers have turned to artificial intelligence to solve this, specifically a method called reinforcement learning, where a computer program learns by trial and error. In a typical online learning scenario, the system would have to send the drones out to fly, make mistakes, and learn from the consequences. But in the real world, such trial-and-error flights are dangerous. A drone that is still learning might fly too close to another, crash, or point its sensors in the wrong direction, causing a loss of critical data. To avoid these risks, a team of researchers from universities in China and Canada proposed a new approach called AERIS. Instead of letting drones learn by flying in the real world, this system learns entirely from a fixed collection of past flight records. It is like a pilot studying a logbook of previous flights to become better, rather than learning by crashing a plane. The goal was to create a system that could take these existing, imperfect flight logs and improve the decision-making of the drones without ever requiring them to take a risky test flight.
The core problem the researchers faced was that while a central computer can see the entire picture of what all the drones are doing, each individual drone only knows its own local history. If a drone tries to change its behavior based only on what it sees locally, it might make a move that seems good for itself but bad for the whole team. Conversely, if the system tries to be too aggressive in correcting the drones' actions based on a central computer's advice, it might suggest moves that are so different from the past data that the system becomes unreliable and unsafe. The researchers found that existing methods often fell into one of two traps: they either stuck too closely to the old, imperfect logs and failed to improve, or they wandered too far into unknown territory, leading to dangerous errors.
To solve this, the team developed a specific algorithm named STAR-CRDT. This method works by having a central computer evaluate potential changes to a drone's flight path using the global information from the past logs. However, it does not blindly accept every suggestion. Instead, it acts as a cautious teacher. It looks for small, safe adjustments to the drone's actions that are likely to improve the overall performance of the fleet. It only accepts these adjustments if they are close enough to the safe, recorded behaviors to be trusted, yet far enough to offer a genuine improvement. This process ensures that the drones learn to be better without ever forgetting the safety constraints that kept them from crashing in the first place. The system is designed so that the heavy thinking happens during a training phase using the old data, and the final result is a set of simple, local rules that each drone can follow on its own in real-time.
When the researchers tested this new framework, the results were significant. In simulations involving multiple drones flying over random paths, the new system improved the overall performance of the fleet by nearly thirty percent compared to the best existing methods. More importantly, it achieved this while drastically reducing the number of dangerous situations. The system increased the success rate of sensing targets by almost five percent and reduced the risk of collisions by more than half. It also managed to keep the communication quality high, ensuring that the drones could talk to each other and the ground users effectively. Perhaps most impressively, the researchers tested the trained drones on completely new maps based on real city streets, which the system had never seen before. Even in these unfamiliar environments, the drones performed better than any other method, successfully navigating complex urban layouts while maintaining safety and service quality.
The study demonstrates that it is possible to create highly effective, safe, and cooperative drone networks without the need for risky real-world experimentation. By learning from fixed records of past flights and using a careful, trust-based method to refine those lessons, the researchers showed that artificial intelligence can be made both powerful and reliable. This approach offers a practical path forward for deploying advanced sensing and communication networks in the real world, where safety and reliability are paramount. The findings suggest that with the right training strategy, autonomous systems can learn to handle complex, dynamic environments effectively, turning the potential chaos of multiple moving drones into a coordinated and efficient operation.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.