Dr Heriberto Cuayahuitl Portilla (University of Lincoln), Dr Mingjun Zhong (University of Aberdeen), Dr Jelena Vlajic (Queen's University Belfast), Dr. Abdalkarim Mohtasib (Zebra Technologies UK Ltd).

Modern Agentic AI integrates Large Language Models (LLMs), Visual-Language Models (VLMs), and Vision-Language-Action Models (VLAs) to enable reasoning, perception, and action in an  environment. Each model requires substantial computational resources and combining them into a unified agentic system further increases these demands, limiting practical deployment in real-world scenarios such as warehouses, our domain of interest. Warehouses themselves are energy-intensive, generate substantial waste, and involve complex logistics. Therefore, optimising warehouse operations while making agentic AI more computationally efficient can (i) reduce carbon footprint and energy costs, (ii) improve operational performance, and (iii) increase the scalability of agentic AI for broader, cost-effective deployments.

The scientific challenge is therefore twofold: (1) to develop compute-efficient machine learning methods that retain high performance, and (2) to optimise operations in a warehouse domain such as fresh-produce. The former will be addressed using techniques such as mixed-precision fine-tuning, distillation, sparse transformers, and neural architecture search with Bayesian optimisation, among others. The latter will integrate LLMs, VLMs, VLAs, digital twins, and simulators with realistic warehouse operational data to provide a testbed for a more energy- and waste-efficient agentic AI.

The proposed topic is important and timely because it addresses urgent sustainability concerns in modern AI while advancing its application to the concrete domain of warehousing. Both the development of compute-efficient agentic AI and its application to warehouse operations aim to contribute to a more sustainable world by enabling scenarios with a lower carbon footprint than current state-of-the-art methods.

This project offers several opportunities: a supportive research environment focused deep learning in core and applied machine learning (ML); collaborating with an industry partner in a warehouse domain; flexibility to balance the dedication effort to ML methods and their application; building on top of state- of-the-art ML/AI tools; and contributing to a more sustainable AI.