LLM-enabled control with dynamic compute offloading of AI modules

• Grant number: 25/01185-1
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: April 01, 2025
• End date: March 31, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Fabio Luciano Verdi
• Grantee: João Vitor Naves Mesa
• Host Institution: Centro de Ciências em Gestão e Tecnologia (CCGT). Universidade Federal de São Carlos (UFSCAR). Campus de Sorocaba. Sorocaba , SP, Brazil
• Company: Universidade Estadual de Campinas (UNICAMP). Faculdade de Engenharia Elétrica e de Computação (FEEC)
• Associated research grant: 21/00199-8 - SMART NEtworks and ServiceS for 2030 (SMARTNESS), AP.PCPE

Abstract

The introduction of large language models (LLMs) has significantly transformed robotics, enabling applications in different areas such as industrial automation, and human-robot interaction. LLMs are increasingly integrated into robotic control and planning systems, from lower-level tasks to higher-level planning. However, adding an LLM to a robot control loop together with other AI workloads, such as image and scene recognition, increases resource usage, overall latency and energy consumption. Dynamic computational offloading strategies allow AI workloads to be better split between the end-devices and different network (edge) locations. This is a scenario that has been explored in recent initiatives like AI-RAN Alliance, where AI edge applications leverage the RAN infrastructure to optimize their performance and enable new use cases. This research project has the ultimate goal to extend a dynamic compute offloading framework with capabilities to support dynamic distribution of AI workloads at the edge and enable new critical applications such as LLM-based control.