Multi-UAV Task Offloading Optimized with DRL and LLMs
A recent study available on arXiv (2605.04436) introduces a hierarchical model for task offloading in dense urban settings, utilizing multiple UAVs in the Internet of Vehicles (IoV). This approach aims to reduce both delay and energy usage by simplifying a challenging non-convex optimization issue. Initially, a sequential distributed optimization technique leveraging Second-Order Cone Programming (SOCP) refines the 3D flight paths of each UAV to enhance network coverage. Additionally, a combined resource scheduling strategy employs Deep Reinforcement Learning (DRL) and Large Language Models (LLMs): the DRL agent manages the primary resource distribution, while the LLM serves as a semantic macro-scheduler to address task imbalances. A reward decoupling mechanism is also introduced to boost overall efficacy.
Key facts
- Paper arXiv:2605.04436 investigates multi-UAV joint base station-assisted IoV task offloading
- System targets dense urban environments
- Optimization problem is decoupled into a hierarchical execution framework
- First stage uses SOCP for 3D flight trajectory optimization
- Second stage uses DRL and LLMs for hybrid resource scheduling
- LLM acts as semantic macro-scheduler to rectify allocation imbalances
- Reward decoupling mechanism is introduced
- Goal is to minimize system delay and energy consumption
Entities
Institutions
- arXiv