AI Takes Flight: New Deep Learning Method Dramatically Improves Drone Data Collection Efficiency

Researchers combine large language models with reinforcement learning to create smarter, more energy-efficient unmanned aerial vehicles for Internet of Things applications

Scientists have developed a groundbreaking approach to drone navigation that combines the power of large language models with reinforcement learning, potentially transforming how unmanned aerial vehicles (UAVs) collect data from distributed sensors in smart cities, environmental monitoring systems, and disaster response scenarios.

The new method, called Large Language Model-empowered Critic-Regularized Decision Transformer (LLM-CRDT), addresses a critical challenge in drone operations: how to efficiently collect data from multiple sensors while maximizing energy efficiency and minimizing flight time. In testing, the approach achieved up to 36.7% higher energy efficiency than current state-of-the-art decision transformer approaches, according to research published by a team at Queen Mary University of London and Southeast University.

The Problem with Current Drone Operations

Traditional approaches to drone navigation fall into two main categories, each with significant limitations. Optimization-based methods require perfect knowledge of communication conditions throughout the entire mission—information that's rarely available in real-world scenarios where wireless environments change unpredictably. Meanwhile, reinforcement learning (RL) has been widely employed to learn UAV trajectory planning and resource allocation policies through interactions with the environment, eliminating the need for prior knowledge of channel state information, but these methods require extensive and potentially risky real-time interactions with the environment.

Drone/ unmanned aerial vehicles (UAV) surveillance for object/ human detection is familiar in large gatherings in the modern cities era, and the demand for autonomous UAV systems continues to grow across industries. Global Drone Market Growth: Revenues soared by double digits throughout 2024, propelled by the rising need for UAVs in professional settings, with over half of new drone registrations going to professional sectors including inspection, search and rescue, and precision agriculture.

A Revolutionary Hybrid Approach

The LLM-CRDT method represents a significant departure from traditional approaches by treating drone navigation as a sequence modeling problem rather than a trial-and-error learning process. The system uses a pre-trained large language model—specifically GPT-2—as its backbone, then fine-tunes it using a parameter-efficient technique called LoRA (Low-Rank Adaptation).

The key innovation lies in the integration of "critic networks" that evaluate the quality of potential actions, enabling the system to learn effective policies from suboptimal training data. This addresses a major limitation of existing decision transformer methods, which are constrained by the quality of their training datasets and cannot exceed the performance of the data used to train them.

Emerging Applications Across Industries

The timing of this breakthrough aligns with rapidly expanding applications of AI-enhanced drones across multiple sectors. Traditional search and rescue methods in wilderness areas can be time-consuming and have limited coverage. Drones offer a faster and more flexible solution, but optimizing their search paths is crucial for effective operations. Recent research has demonstrated deep reinforcement learning to create efficient search paths for drones in wilderness environments, showing promising results in emergency response scenarios.

In the medical sector, researchers have developed a learning-based coordinated unmanned aerial vehicle–unmanned ground vehicle (UAV–UGV) (CUU) framework, currently unavoidable use, with a transfer learning algorithm for medical waste transportation, highlighting the versatility of AI-driven drone systems.

The Large Language Model Advantage

The integration of large language models with UAV systems represents a rapidly emerging field. Large Language Models (LLMs), a key component of AI, exhibit remarkable learning and adaptation capabilities within deployed environments, demonstrating an evolving form of intelligence with the potential to approach human-level proficiency. Recent work has explored how LLM integration expands the scope of existing UAV applications, enabling autonomous data processing, improved decision-making, and faster response times in emergency scenarios like disaster response and network restoration.

The majority of studies explore how LLMs can enable natural language control of drones, assist in planning complex missions, improve decision-making in dynamic environments, and create more autonomous and intelligent aerial agents. However, research reveals a strong focus on large language models (LLMs) and their integration with robotics and unmanned aerial vehicles (UAVs), with academic research showing 40.4% emphasis on theoretical modeling.

Technical Innovation and Performance

The LLM-CRDT system addresses several technical challenges simultaneously. First, it solves the resource allocation problem optimally using linear programming techniques, reducing computational complexity while ensuring efficient use of communication resources. Second, it leverages the pre-trained knowledge in large language models to reduce the amount of training data required—a significant advantage given the cost and difficulty of collecting high-quality drone trajectory data.

The system was tested against multiple benchmark methods, including online reinforcement learning approaches like TD3 and SAC, as well as offline methods such as behavior cloning and other decision transformer variants. The results consistently showed superior performance across different scenarios, with the LLM-CRDT method achieving higher energy efficiency while maintaining robust performance even when trained on suboptimal data.

Real-World Applications and Future Directions

The implications extend far beyond academic research. A public safety Unmanned Aerial Vehicle (UAV) enhances situational awareness in emergency response. Its agility and ability to optimize mobility and establish Line-of-Sight (LoS) communication make it increasingly vital for managing emergencies such as disaster response, search and rescue, and wildfire monitoring.

In disaster scenarios, establishing robust emergency communication networks is critical, and unmanned aerial vehicles (UAVs) offer a promising solution to rapidly restore connectivity. The new framework could enable more efficient deployment of drone swarms for emergency communications, with multi-agent reinforcement learning (MARL) and large language models (LLMs) to jointly optimize UAV agents toward achieving optimal networking performance.

Looking Ahead

As the field continues to evolve, researchers are exploring even more sophisticated applications. The system combines deep reinforcement learning (RL) in simulation with data collected in the physical world. Swift competed against three human champions, including the world champions of two international leagues, in real-world head-to-head races, demonstrating that AI systems can now match or exceed human performance in complex physical tasks.

The convergence of large language models and drone technology represents a significant step toward fully autonomous aerial systems that can adapt to complex, dynamic environments while operating efficiently and safely. As these technologies mature, they promise to revolutionize industries ranging from environmental monitoring and disaster response to smart city management and precision agriculture.

Sources:

1. Chen, Z., Wang, J., Shin, H., & Nallanathan, A. (2025). Large Language Model-Empowered Decision Transformer for UAV-Enabled Data Collection. arXiv preprint arXiv:2509.13934. https://arxiv.org/html/2509.13934 [2509.13934] Large Language Model-Empowered Decision Transformer for UAV-Enabled Data Collection

2. Scientific Reports. (2025). Drone-assisted adaptive object detection and privacy-preserving surveillance in smart cities using whale-optimized deep reinforcement learning techniques. Nature. https://www.nature.com/articles/s41598-025-94796-3

3. DSLRPros. (2025). Next-Gen UAVs: Cutting-Edge Drone Innovations to Watch in 2025. https://www.dslrpros.com/blogs/rescue-drones/next-gen-uavs-cutting-edge-drone-innovations-to-watch-in-2025

4. Sharma, D. D., & Lin, J. (2024). Secure learning-based coordinated UAV–UGV framework design for medical waste transportation. Frontiers in Remote Sensing, 5:1351703. https://www.frontiersin.org/journals/remote-sensing/articles/10.3389/frsen.2024.1351703/full

5. Ewers, J.-H., Anderson, D., & Thomson, D. (2025). Deep reinforcement learning for time-critical wilderness search and rescue using drones. Frontiers in Robotics and AI, 11:1527095. https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2024.1527095/full

6. Medaiyese, O., et al. (2024). UAV Detection Using Reinforcement Learning. PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC10975760/

7. Kaufmann, E., et al. (2023). Champion-level drone racing using deep reinforcement learning. Nature. https://www.nature.com/articles/s41586-023-06419-4

8. Frontiers in Neurorobotics. (2025). Unmanned aerial vehicle based multi-person detection via deep neural network models. https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2025.1582995/full

9. Kong, X., Zhou, Y., Li, Z., & Wang, S. (2024). Multi-UAV simultaneous target assignment and path planning based on deep reinforcement learning. Frontiers in Neurorobotics, 17:1302898. https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2023.1302898/full

10. Javaid, S., et al. (2024). Large Language Models for UAVs: Current State and Pathways to the Future. arXiv preprint arXiv:2405.01745. https://arxiv.org/html/2405.01745v1

11. Emami, Y., et al. (2025). From Prompts to Protection: Large Language Model-Enabled In-Context Learning for Smart Public Safety UAV. arXiv preprint arXiv:2506.02649. https://arxiv.org/html/2506.02649v1

12. Chen, Y., et al. (2025). When Large Language Models Meet UAVs: How Far Are We? arXiv preprint arXiv:2509.12795. https://arxiv.org/html/2509.12795v1

13. Quantum Zeitgeist. (2025). Large Language Models And UAVs: Study Of 74 Papers Reveals 9 Task Types And 40.4% Emphasis On Theoretical Modeling. https://quantumzeitgeist.com/40-4-percent-models-large-language-uavs-papers-reveals-task-types-emphasis/

14. MRLMN Framework. (2025). Scalable UAV Multi-Hop Networking via Multi-Agent Reinforcement Learning with Large Language Models. arXiv preprint arXiv:2505.08448. https://arxiv.org/html/2505.08448v1

15. UAVs Meet LLMs. (2025). Overviews and Perspectives Toward Agentic Low-Altitude Mobility. arXiv preprint arXiv:2501.02341. https://arxiv.org/html/2501.02341v1