Artificial General Intelligence (AGI)-Native Wireless Systems: Digital Twins and World Models for Beyond 6G Networks

Abstract

Building next-generation wireless systems that can reliably support physical artificial intelligence (AI) agents (e.g., robots, autonomous vehicles, etc.) requires advanced levels of intelligence beyond today's state-of-art. On the one hand, the 6G vision of AI-native networks typically relies on standard AI methods (e.g., neural networks) that perform poorly in non-stationary, real-world environments. On the other hand, physical AI agents still lack the capability to generalize and adapt in unforeseen scenarios that appear in real-world settings. As a result, today's AI-native wireless systems and, in turn, their physical AI agents remain far from being autonomous and fall short in terms of quality-of-service. To address this limitation, this dissertation aims to revisit and redefine the concept of AI-native wireless systems, equipping them with common sense capabilities necessary to transform them into artificial general intelligence (AGI)-native systems. This transformation promises to revolutionize wireless systems by enabling unprecedented levels of cognitive wireless intelligence. Notably, such intelligence provides networks with the reasoning, planning, and complex inference needed to operate in dynamic, real-world environments. This envisioned new generation of networks is driven by a cognitive brain architecture that is founded on three components: A perception module, a world model, and an action-planning component. Towards realizing these components, first, we show how the perception module can be built through abstracting real-world elements into generalizable representations. Then, these representations are used to form a world model, founded on principles of causality and hyper-dimensional computing. Subsequently, we design intent-driven and objective-driven planning methods that can maneuver the network to take its actions. Central to this architecture, world models offer a structured approach for mirroring the physical world into a digital counterpart over the network. Cheif among these counterparts are the digital twins (DTs) of physical AI agents. With this interconnection to world models, DTs offer a gateway to instill common sense from the network directly into these agents. Nevertheless, enabling such solution requires addressing novel wireless challenges. Chief among those challenges is preserving the synchronization of DTs and world models with the physical world. To address this challenge, we propose a rigorous decentralized framework that decomposes these world models and their DTs over the network edge. In particular, we pose an optimization problem that aims to minimize the synchronization delays of smaller-scale world models and their associated DTs at the edge, while ensuring their interoperability in terms of association and resource allocation. To solve this problem, we propose an optimal transport theory algorithm that ensures the optimal average synchronization time of the world models, while satisfying the synchronization intensity requirements of the DTs. Results show that synchronization delays can be reduced up to 25 % in comparison to the standard signal-to noise ratio (SNR) association benchmark. Accordingly, the decentralized DTs and world models are then leveraged to drive reasoning back into the physical AI agents in the real world. This reasoning allows the physical AI agents to generalize when facing unforeseen scenarios, thereby enabling a revolutionary test-time scaling law for physical AI agents. In particular, this novel scaling law builds on the first principle of active inference and extends action selection to incorporate inference-driven reasoning that scales the feed-forward policy in unforeseen scenarios. Nevertheless, this decision-making process is formulated as a partially observable Markov decision process (POMDP) that renders an intractable inference problem. To obtain a tractable solution for this POMDP, this problem is solved through a variational Bayesian approach that unifies perception, planning, action, and learning under the minimization of variational and expected free energy. Results showcase how the proposed framework yields AI agents that can act, reason, learn, and maintain generalizable performance in dynamic environments. As continual learning (CL) abilities emerge while updating the DT models with these unforeseen scenarios, we further design a deep CL solution to enable synchronized model updates for physical AI agents. In particular, we propose a novel CL solution that preserves the accuracy and synchronization of the evolving DTs at the edge. To limit the de-synchronization gap arising during the DT model update, this update process is posed a dual objective optimization problem whose goal is to jointly minimize the loss function over all encountered historical episodes and the corresponding de-synchronization time. As the de-synchronization time continues to increase over sequential historical episodes, an elastic weight consolidation (EWC) technique that continually regularizes the DT history is proposed to limit de-synchronization time. Furthermore, to address the plasticity-stability tradeoff accompanying the progressive growth of the EWC regularization terms, a modified EWC method that considers fair execution between the historical episodes of the DTs is adopted. Simulation results show that the proposed solution can achieve an accuracy of 90% while guaranteeing a minimal de-synchronization time. Therefore, this dissertation is expected to shape the future of DTs and world models as enablers of AGI over future networks. Ultimately, this dissertation serves as a blueprint to drive the next generation of wireless networks and its autonomous physical AI agents in the 6G and beyond era.