Problem

Current artificial intelligence (AI) systems, including deep learning and generative AI, often fail when integrated into physical systems like robots or vehicles. This limitation stems from their inability to maintain reliable world models for long-horizon planning under uncertainty, generalize to unforeseen scenarios, and adapt to dynamic real-world conditions. Furthermore, existing wireless networks are primarily designed as passive data transport layers, optimizing for throughput and latency, which makes them incapable of supporting the real-time coordination and shared spatiotemporal context essential for physical AI. Key challenges include the lack of reliable self-models and environmental models in physical AI, limited causal reasoning, difficulty maintaining accurate beliefs over extended periods, and significant domain shifts from simulation to real-world deployment. Wireless networks exacerbate these issues by not prioritizing transmissions based on "cognitive value," lacking precise synchronization for closed-loop physical actions, failing to align heterogeneous data representations, and having no "theory of mind" capabilities to understand agent intentions.

Method

The paper proposes the Network of Holonic Digital Twins (HDT-Nets) framework, which reimagines wireless networks as active orchestrators of physical intelligence. In this framework, each Holonic Digital Twin (HDT) functions as both an autonomous cognitive agent reasoning locally and a cooperative component of larger, collectively intelligent units. The HDT-Net is built upon four fundamental pillars:

1. Causal Markov Blankets (MBs): These dynamically partition cyber-physical networks into autonomous AI agents, defining their boundaries and determining which agents must coordinate and what information to exchange based on causal relevance. MBs allow for efficient resource allocation by identifying minimal variables for coordination and enabling counterfactual reasoning.
2. Category Theory: This mathematical framework ensures compositional consistency and semantic interoperability among heterogeneous HDTs with different sensor modalities and world model representations. It formalizes how abstraction levels compose (e.g., via functors for physical-to-virtual mapping, abstraction, and theory of mind) and how beliefs are coherently translated across incompatible semantic languages (via natural transformations).
3. Active Inference: This principle unifies perception, action, and learning within each HDT by minimizing expected free energy. It drives belief updates about hidden states, selects actions that balance goal achievement and uncertainty reduction, and continuously refines the causal generative model. Active inference also formalizes "cognitive value" of information, prioritizing transmissions that significantly reduce a receiver's expected free energy.
4. Integrated Information Theory (IIT): This theory introduces spatiotemporal integrated information (Φ) as a quantifiable metric to measure when collective intelligence genuinely emerges from the network of coordinating HDTs, surpassing the sum of individual contributions. Φ helps in guiding the allocation of sensing, communication, computing, and control resources to actively cultivate and maximize collective intelligence growth over time.

Results

The HDT-Net framework transforms wireless networks from passive data processors into active cognitive systems capable of reasoning about their environment, understanding uncertainty, coordinating through meaning, and dynamically adapting their integration. This architecture enables several crucial capabilities for physical AI at scale. Firstly, it provides a shared spatiotemporal context, ensuring that distributed agents perceive, decide, and act within a unified physical reference frame. Secondly, it facilitates safety-critical coordination by evaluating the cognitive value of every transmission based on its impact on the receiver's subsequent physical actions, rather than just raw data volume. Thirdly, HDT-Nets allow for seamless navigation of the physical world, maintaining collective intelligence even as agents move, missions evolve, and infrastructure degrades. The framework also leads to self-organizing network topologies, where coordination links are established or released based on the mutual information carried versus communication cost, thereby optimizing for epistemic objectives rather than just throughput.

Implications

The paper identifies several critical open research challenges for the practical deployment and advancement of HDT-Nets. These include developing new models for cognitive traffic generation and belief-centric communication metrics that go beyond traditional bit error rates to quantify belief fidelity and cognitive capacity. Addressing latency is crucial, as delayed or causally misattributed belief updates can lead to safety-critical failures, necessitating new "world model belief age" metrics. Designing adaptive network topologies that optimize collective epistemic uncertainty under wireless resource constraints is another key challenge. Furthermore, co-designing belief delivery, inference scheduling, and action commitment across wireless infrastructure is required to ensure a shared temporal context for physically coupled swarms with deterministic timing guarantees. Scalable computation of spatiotemporal integrated information (Φ) is essential for large networks, requiring local approximations and hierarchical decomposition. Finally, ensuring adversarial resilience and trust in holonic coordination, where belief integrity at the semantic level is protected against compromised agents, presents a significant security challenge. The transition to 6G networks, as envisioned by HDT-Net, signifies a shift towards networks that understand uncertainty, coordinate through meaning, and collectively grow smarter over time.