Robust Foundation Models Empowered RAN Intelligence for Reliable Embodied Robot Scenarios

longyu zhou, Yijing Liu, Tony Quek, Wenbo Jiang

Keywords: Robustness, Reliability, foundation model, O-RAN

TL;DR: A robust FM-empowered O-RAN framework is proposed to enable secure, real-time, and reliable robot cooperation for parcel sorting, achieving high accuracy, low latency, and enhanced operational safety.

Abstract: The rapid development of large-scale AI has made intelligent robots increasingly viable for applications such as warehouse parcel sorting. Coupled with advances in mobile communication, robots can now cooperate efficiently, yet conventional AI-based solutions often face low resource utilization and limited robustness, hindering both sorting accuracy and handling efficiency. To address this, we propose a robust Foundation Model (FM)-empowered O-RAN framework that enables secure, robust, and real-time robot cooperation. An adaptive FM-splitting algorithm decomposes tasks into sequential sub-missions to improve sorting accuracy, while robustness training ensures resilience to environmental variations. Additionally, a cooperative path planning algorithm optimizes the number of active robots, reducing handling latency and energy consumption. Experiments demonstrate stable GPU utilization, up to 90% sorting accuracy, a 13.9% reduction in latency, and enhanced operational safety compared with conventional FM-based approaches.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 6871