DRASTIC: A Dynamic Resource Allocation Framework over 6G Network Slicing in Task-aware Closed-Loop Tactile Internet Applications

This work proposes a novel learning driven bandwidth optimization framework called DRASTIC (Dynamic Resource Allocation for Slicing in Task aware Closed loop tactile Internet applications). The proposed framework dynamically allocates resources among network slices supporting both enhanced Mobile Broadband (eMBB) and high reliable low latency communication (HRLLC) users. The algorithm ensures queue stability and meets delay targets with high probability under a Markov-modulated Poisson traffic, exploiting a Lyapunov guided advantage actor critic reinforcement learning technique. The proposed network model includes an open-loop eMBB queue whose arrival and departure are mainly driven by throughput demand, as well as a closed loop HRLLC queue that captures feedback and task execution effects. A task execution dependent dexterity index adjusts the effective arrival rate, creating a feedback aware interaction between the network and the task. A probabilistic delay constraint is incorporated into the objective via Lagrangian relaxation, yielding a min_max optimization framework that enforces latency guarantees while maximizing throughput for both types of users. Simulation results demonstrate that the proposed framework meets diverse Quality of Service (QoS) requirements, maintains queue stability under dynamic wireless and robotic task variation conditions, and outperforms other approaches.

💡 Research Summary

The paper introduces DRASTIC, a learning‑driven dynamic resource allocation framework designed for 6G network slicing that simultaneously serves enhanced Mobile Broadband (eMBB) and high‑reliability low‑latency communication (HRLLC) traffic in tactile‑Internet applications. Recognizing that tactile‑Internet traffic is inherently closed‑loop—where the generation rate of control packets depends on the robot’s task difficulty—the authors model HRLLC arrivals using a two‑state Markov‑modulated Poisson process (MMPP). A scalar dexterity index (DXI) reported by the robot modulates the instantaneous arrival intensity, reducing packet generation when the task is hard and increasing it when the task is easy. This captures the feedback‑aware interaction between communication and control that traditional open‑loop models ignore.

The system consists of a single gNB with total bandwidth B divided into K physical resource blocks (PRBs). Separate queues are maintained for HRLLC (F_i) and eMBB (G_i) users. The achievable rate on each PRB follows the Shannon formula with Rayleigh fading, and the total service rate for a user is the sum over allocated PRBs. The HRLLC queue dynamics incorporate the DXI‑adjusted MMPP arrivals, while the eMBB queue follows a fixed‑rate Poisson arrival process.

The objective is to maximize long‑term average data rates for both slices while guaranteeing a probabilistic delay constraint for each HRLLC user: the probability that end‑to‑end delay exceeds a deadline D_max must be less than 1 − χ_h. Directly optimizing this tail probability is intractable, so the authors introduce an exponential surrogate function that embeds the queue backlog, service rate, packet size, and reliability target into a single penalty term. By applying Lagrangian relaxation, the problem becomes a min‑max formulation where a virtual penalty weight adaptively penalizes delay violations.

To ensure queue stability, a Lyapunov drift‑plus‑penalty framework is employed. The drift term pushes the system toward smaller backlogs, while the penalty term (scaled by a tunable parameter V) balances throughput against delay reliability. The per‑slot PRB allocation decision is cast as a Markov Decision Process (MDP). The authors solve the MDP with an Advantage Actor‑Critic (A2C) reinforcement learning algorithm. The actor network outputs a continuous allocation vector, which is subsequently quantized to integer PRB counts; the critic estimates the state‑value function to compute the advantage used for policy updates. This approach yields sample‑efficient learning and can handle the high‑dimensional state space comprising queue lengths, channel quality indicators (CQI), HARQ feedback, and the dexterity index.

A key contribution is the integration of the framework with the Open RAN (O‑RAN) architecture. The non‑real‑time RIC (timescales > 1 s) performs long‑term analytics, determines slice‑level objectives, and disseminates updated policy parameters via the A1 interface. The near‑real‑time RIC (10–100 ms) hosts DRASTIC as an xApp, receives live measurements from the O‑DU through the E2 interface, runs lightweight inference using the learned policy, and returns per‑slot PRB allocation decisions to the DU for execution. This design respects O‑RAN timing constraints while enabling closed‑loop, slice‑aware scheduling that adapts to bursty HRLLC traffic and time‑varying channel conditions.

Extensive simulations evaluate DRASTIC under various MMPP transition rates (α, β) and dexterity index dynamics, comparing against Round‑Robin and Proportional‑Fair baselines. Results show that DRASTIC maintains queue stability, achieves HRLLC delay reliability exceeding 99.9 % (χ_h ≥ 0.999), and incurs only marginal eMBB throughput loss. The probability of HRLLC delay violation is reduced by more than a factor of three relative to the baselines, confirming the effectiveness of the Lyapunov‑A2C‑Lagrangian combination.

In summary, DRASTIC provides a unified, AI‑native solution for 6G network slicing in task‑aware tactile‑Internet scenarios. By jointly modeling traffic burstiness, task‑dependent arrival modulation, probabilistic latency constraints, and by embedding the solution within the O‑RAN control hierarchy, the framework advances the state‑of‑the‑art in delivering ultra‑reliable low‑latency services alongside high‑throughput eMBB traffic.