Temporal Windows of Integration for Multisensory Wireless Systems as Enablers of Physical AI

Physical artificial intelligence (AI) refers to the AI that interacts with the physical world in real time. Similar to multisensory perception, Physical AI makes decisions based on multimodal updates from sensors and devices. Physical AI thus operates with a finite spatial footprint of its sensory tributaries. The multimodal updates traverse heterogeneous and unreliable paths, involving wireless links. Throughput or latency guarantees do not ensure correct decision-making, as misaligned, misordered, or stale inputs still yield wrong inferences. Preserving decision-time coherence hinges on three timing primitives at the network-application interface: (i) simultaneity, a short coincidence window that groups measurements as co-temporal, (ii) causality, path-wise delivery that never lets a consequence precede its precursor, and (iii) usefulness, a validity horizon that drops information too stale to influence the current action. In this work, we focus on usefulness and adopt temporal window of integration (TWI)-Causality: the TWI enforces decision-time usefulness by assuming path-wise causal consistency and cross-path simultaneity are handled upstream. We model end-to-end path delay as the sum of sensing/propagation, computation, and access/transmission latencies, and formulate network design as minimizing the validity horizon under a delivery reliability constraint. In effect, this calibrates delay-reliability budgets for a timing-aware system operating over sensors within a finite spatial footprint. The joint choice of horizon and per-path reliability is cast as a convex optimization problem, solved to global optimality to obtain the minimal horizon and per-path allocation of reliability. This is compared favourably to a benchmark based on uniform-after-threshold allocation. Overall, this study contributes to timing-aware Physical AI in next-generation networks.

💡 Research Summary

The paper addresses a fundamental timing problem for Physical Artificial Intelligence (Physical AI), which must fuse multimodal sensor updates in real time to act as an intelligent organism within a finite spatial footprint. While traditional communication metrics such as throughput or latency guarantees are insufficient, the authors focus on the “usefulness” primitive: a validity horizon (Δ) that discards any update older than a prescribed staleness bound. They assume that simultaneity (co‑temporal grouping) and causality (path‑wise ordering) are already enforced upstream via timestamps and hybrid logical clocks (HLCs). Consequently, the network design problem reduces to determining the smallest possible Δ that still satisfies a global reliability target for all admissible causal paths delivering updates to a base‑station/edge processor.

To this end, the authors construct a detailed end‑to‑end latency model for each causal path. The model aggregates three components: (i) event‑to‑sensor propagation delay (distance‑dependent), (ii) bounded computation latency at each hop (sensor, relay, edge), and (iii) access and transmission latency under a slotted‑frame MAC with retry limits. Multi‑hop relays are modeled using a “truncated‑attempt” approach: intermediate hops employ grant‑free access, while the final hop uses grant‑based scheduling. Closed‑form expressions for packet‑drop probability and the probability mass function (PMF) of total path delay are derived.

Reliability is defined as the probability that a path’s delay does not exceed Δ. The system must allocate a per‑path “violation budget” ε_i such that the sum of ε_i across all paths respects the global reliability requirement (e.g., 99.9 % success). Because the exact delay distributions are complex, the authors employ a distribution‑agnostic bound based on the first two moments (mean μ_i and variance σ_i²) using Cantelli’s one‑sided Chebyshev inequality. This yields a linear constraint of the form ε_i ≥ (σ_i²)/(σ_i² + (Δ – μ_i)²), which is convex in the decision variables.

The overall optimization problem is: minimize Δ subject to (a) the Cantelli‑derived per‑path constraints, (b) Σ ε_i ≤ 1 – target reliability, and (c) ε_i ≥ 0. The problem is convex; the authors solve it via a bisection search on Δ, with each candidate Δ checked for feasibility by solving a simple linear program for the ε_i variables. The solution provides both the minimal validity horizon and the optimal allocation of reliability budgets across heterogeneous paths.

Numerical experiments compare the optimal non‑uniform allocation against a baseline “uniform‑after‑threshold” (water‑filling) scheme, where all paths receive equal ε_i up to a threshold and excess reliability is distributed uniformly. The optimal scheme consistently reduces Δ by 20–30 % in scenarios with heterogeneous path delays, demonstrating the benefit of shaping reliability according to each path’s delay statistics.

Sensitivity analyses explore how system parameters affect the optimal Δ and ε_i: larger cell radii increase propagation delay, inflating Δ; shorter frame durations mitigate MAC‑induced latency; deeper causal paths increase the impact of retransmission limits; and tighter global reliability targets force smaller ε_i, thereby enlarging Δ. These insights translate into concrete design guidelines for 6G perceptive wireless networks, suggesting where to invest in faster links, tighter synchronization, or more aggressive MAC configurations.

Finally, the authors highlight practical implications: the per‑path reliability budgets can serve as weighting factors in multimodal fusion algorithms, improve inference accuracy, and enable temporal forensics (bias‑aware analysis) by exposing which sensor streams contributed within the validity horizon. The work thus delivers a network‑centric methodology for timing‑aware Physical AI, bridging the gap between human‑inspired multisensory perception and engineered wireless systems.