Safe and Economical UAV Trajectory Planning in Low-Altitude Airspace: A Hybrid DRL-LLM Approach with Compliance Awareness

TL;DR

The paper tackles UAV trajectory planning in dense low-altitude urban airspace where safety, regulatory compliance, and energy efficiency are all critical. It introduces a hybrid framework that combines soft actor-critic DRL with large language model reasoning, formulated as a POMDP with a composite reward to encourage robust obstacle avoidance, compliance, and economical operation. The LLM provides semantic, rule-aware guidance during training when obstacles are detected, while the deployment phase relies on the trained SAC policy to ensure real-time performance; computational complexity accounts for occasional LLM inferences during training. Experimental results demonstrate that the approach improves data collection rate, reduces collisions and regulatory violations to near zero, achieves 100% successful landings, and lowers energy consumption relative to baselines. This work offers a practical pathway to scalable, compliant, and energy-efficient UAV data collection in the evolving low-altitude economy.

Abstract

The rapid growth of the low-altitude economy has driven the widespread adoption of unmanned aerial vehicles (UAVs). This growing deployment presents new challenges for UAV trajectory planning in complex urban environments. However, existing studies often overlook key factors, such as urban airspace constraints and economic efficiency, which are essential in low-altitude economy contexts. Deep reinforcement learning (DRL) is regarded as a promising solution to these issues, while its practical adoption remains limited by low learning efficiency. To overcome this limitation, we propose a novel UAV trajectory planning framework that combines DRL with large language model (LLM) reasoning to enable safe, compliant, and economically viable path planning. Experimental results demonstrate that our method significantly outperforms existing baselines across multiple metrics, including data collection rate, collision avoidance, successful landing, regulatory compliance, and energy efficiency. These results validate the effectiveness of our approach in addressing UAV trajectory planning key challenges under constraints of the low-altitude economy networking.

Figures (14)

• Figure 1: System model of low-altitude UAV trajectory planning. The DCU collects data from GE while avoiding static obstacles (BZs, NFZs, RZs) and dynamic obstacles (OUs) under energy and compliance constraints.
• Figure 2: The proposed algorithm integrates an LLM into SAC for conditional action selection. The LLM is invoked to generate actions when obstacles are detected; otherwise, actions are selected by the SAC policy network. The chosen actions interact with the environment, and transitions are stored to train the critic networks.
• Figure 3: CoT-based prompt generator used for UAV control. Part (a) defines the task and observation, (b) describes the reasoning principles, and (c) specifies the action format and compact JSON output.
• Figure 4: Training reward curves of the proposed algorithm and all baseline methods over 4000 episodes, including SAC, PPO, DDPG, CPO, Theile et al. theile2024equivariant, Wang et al. wang2023ensuring, SAC+Heuristic, and SAC+LLM variants with different invocation probabilities.
• Figure 5: The DCR, CR, and ECR about baselines and our algorithm when $N_{\mathrm{OU}}$ increases. The proposed algorithm consistently achieves the highest DCR with almost zero CR and lowest ECR, demonstrating superior efficiency and safety.