Title

Exploring Edge Physical Intelligence: Extreme Quantization, Limitless Memory, and Rapid Evolution

Abstract

Physical intelligence, as applied to edge systems such as robotics and autonomous driving, imposes increasingly stringent requirements on intelligent models. These models must not only comprehend and describe the world but also act effectively within it. To address the critical challenges faced by current intelligent models during edge deployment, e.g., limited computational resources, high storage consumption, inadequate inference capabilities, and poor physical adaptability, we undertake a series of novel explorations. Our objective is to develop truly embodied and adaptive edge physical intelligence. Specifically, to mitigate the substantial computational and deployment costs of intelligent models on edge devices, we introduce a policy-adaptive quantization technique. This approach effectively accelerates computation and reduces resource consumption, enabling models to operate efficiently on constrained hardware while maintaining high performance. Considering that edge devices struggle to meet the exponentially growing storage demands of intelligent models, we design and integrate an AI+SSD-based "infinite" memory mechanism to effectively support ultra-long context inference. Furthermore, to overcome the limitations of insufficient cognitive and inference capabilities in edge intelligent models, we propose a reinforcement learning-based "extreme cognitive learning" framework. This framework imbues models with meta-cognitive abilities such as self-reflection, evaluation, and control, thereby significantly enhancing their decision-making performance. Moreover, to tackle the challenges of low learning efficiency and slow adaptation to novel physical environments in physical intelligent models, we introduce a world model-driven "rapid evolution strategy." This strategy enables iterative interaction between simulated and real environments, leading physical intelligent models to efficiently acquire and optimize physical skills.

Biography

Song Guo is a Chair Professor in the Department of Computer Science and Engineering at the Hong Kong University of Science and Technology. Prof. Guo made fundamental and pioneering contributions to the development of edge AI and machine learning systems. He has published many papers in top venues with wide impact in these areas and has been consistently recognized as a Clarivate Highly Cited Researcher. He is the recipient of the IEEE 2024 Edward J. McCluskey Technical Achievement Award, and over a dozen Best Paper Awards from IEEE/ACM. Prof. Guo is the Editor-in-Chief of IEEE Transactions on Cloud Computing. He has served on the IEEE Fellow Evaluation Committee for both the Computer and Communications Societies. He has also served as organizing and technical committee chair for many IEEE/ACM conferences and workshops. Prof. Guo is a Fellow of the Canadian Academy of Engineering, a Member of Academia Europaea, and a Fellow of the IEEE.