Researchers in China have proposed a new approach to one of the thorniest problems in autonomous mobility: deciding which vehicle in a self-driving delivery fleet should serve which order, and when. The answer, they suggest, may lie in the same large language model (LLM) technology that powers conversational chatbots.
The framework, named Predictive-context Large Language Model Scheduler (P-LLM), was developed by a team led by Jiaxin Tan and corresponding author Xiaohui Huang at East China Jiaotong University in Nanchang, according to a report by Gasgoo, a Chinese automotive news outlet. The research was published in the journal Applied Intelligence.
Instead of relying solely on hand-crafted optimization rules or reinforcement learning agents that require weeks of training, P-LLM lets the LLM reason directly about fleet status, real-time order flow, and predicted demand that has not yet materialized. The scheduler then proposes vehicle-order matches that balance immediate profit against the expected geographic distribution of future demand.
Why traditional methods fall short
Classic optimization algorithms work well when the environment is static and rules are clear, but real urban environments are anything but static. Traffic conditions change rapidly, and order demand can surge or dwindle without warning. A decision that looks optimal in the moment could prove disastrous five minutes later if it strands a vehicle in a zone where demand is about to spike.
Reinforcement learning, a modern alternative, learns policies through trial and error, but the authors note that in noisy, non-stationary urban traffic, these methods still face serious challenges in training stability and convergence.
The P-LLM framework addresses these issues by having the LLM reason over a broad context that includes both current conditions and predictive insights. This allows the system to anticipate where demand is likely to emerge, rather than merely reacting to what has already happened.
The research is part of a broader trend in China's EV and autonomous driving sectors, where AI models are increasingly being applied to operational challenges beyond the vehicle itself. While the study is academic, it points to a future where LLMs could play a central role in managing fleets of autonomous delivery vehicles in smart cities.
The team behind P-LLM includes Nan Jiang, Xuebo Cheng, Liyan Xiong, and Ni Liu, according to the Gasgoo report. The study was published in Applied Intelligence, a peer-reviewed journal covering artificial intelligence and its applications.
