Title: AUGUSTE: A Learning-Based dApp for Predictive URLLC Scheduling

Abstract

The drive for Ultra-Reliable and Low Latency Communications (URLLC) served as a primary catalyst for the development of 5G technology. According to 3GPP, 5G aims to deliver latency targets between 1 and 10 milliseconds for critical use cases, including industrial automation, Vehicle-To-Everything (V2X) connectivity, tactical edge networking, and the control of unmanned systems. However, years into the deployment of real-world 5G Time Division Duplexing (TDD) networks, median Uplink (UL) round-trip times remain in the 50–70 millisecond range. This discrepancy is largely attributed to the Scheduling Request (SR) protocol, which User Equipment (UE) must execute prior to transmitting UL data.

Current solutions, predominantly Configured Grant (CG) scheduling, mitigate this overhead only for strictly periodic traffic and demand cross-layer synchronization, factors that have constrained their widespread adoption. To address these limitations, we introduce AUGUSTE (Anticipatory Uplink Grants for URLLC via Self-Adapting Temporal Estimation). This is a learning-driven Medium Access Control (MAC) scheduling framework that integrates online Machine Learning (ML) models directly into the UL scheduler. By predicting packet arrivals, the system proactively allocates resources before a user initiates an SR.

AUGUSTE utilizes an adaptive state machine that toggles between two modes: a learning phase, which gathers unbiased arrival statistics, and a confident phase, which leverages these predictions to schedule transmissions only when traffic is anticipated. We assessed the performance of AUGUSTE on a live 5G testbed utilizing OpenAirInterface, testing it against three distinct URLLC traffic scenarios: request-response interactions, edge inference for machine learning, and periodic autonomous reporting. The results demonstrate that AUGUSTE achieves the optimal balance in the latency-overhead trade-off. It matches the median Round Trip Time (RTT) of always-on scheduling—approximately 10 ms, which is half the 20 ms baseline of SR-based methods—while incurring only one-tenth of the resource cost, with an overhead of just 7–10 percent.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC