On-Device Machine Learning Model Adaptation for 6G Communications: From Standardization to Over-the-Air Prototyping Experiments Based on Beam Prediction

Adapting pretrained artificial intelligence and machine learning (AI/ML) models for on-device inference across diverse environments is challenging due to the extensive offline data collection and training required. This highlights the necessity for online model training or adaptation, where devices locally process training data and perform adaptation via on-device engines. Such capabilities and technologies are essential for various wireless AI/ML features requiring on-device inference to be considered in the upcoming Third Generation Partnership Project (3GPP) 6G standardization works. For battery-powered devices, model adaptation must be efficiently executed in real-time with well-designed timelines, as environmental conditions can change rapidly and vary significantly across locations. In this paper, we present results from over-the-air (OTA) prototyping experiments using handset devices to demonstrate the practical effectiveness of real-time on-device adaptation. We use AI/ML based beam prediction, as outlined by 3GPP 5G-Advanced standards, as an example. Our OTA results show that on-device model adaptation can be completed within 30 seconds in an unfamiliar environment to achieve benchmark performance, strongly supporting the feasibility of real-time on-device model adaptation in 6G communications. The findings and methodologies discussed in this paper can be extended to benefit various other wireless and non-wireless real-world scenarios.