SLED: Sensor SLicing from Edge to Datacenter

Alan Subedi, Avhishek Biswas, Apala Pramanik, Amlan Balabantaray, Mikhail Titov, Santosh Pitla, Ozgur Kilic, Mehmet C. VuranarXiv 2026 · arXiv

TL;DR

SLED (Sensor SLicing from Edge to Datacenter) is a novel framework that brings fine-grained network slicing down to the individual sensor level within a single device. By using an AI-driven "agentic coding" workflow, the researchers implemented deterministic resource isolation in Open-RAN stacks, ensuring that critical telemetry isn't "starved" by bulk data transfers—even when they share the same wireless hardware.

What Problem Does This Paper Address?

In the world of 5G and 6G, "network slicing" is a well-known concept used to give different devices (like a phone vs. an autonomous car) different priorities. However, current systems typically slice at the device level.

Imagine a sophisticated field robot equipped with high-res cameras, LiDAR, and critical control telemetry. Under the current standard, all these sensors share one "slice" assigned to the robot. If the robot starts a massive LiDAR data dump, that bulk transfer can create internal contention, causing massive delays for the tiny but time-sensitive control commands. This "intra-device" interference can lead to system failures or degraded performance in high-stakes AI-IoT applications.

Methods

The authors introduce Sensor Slicing (SL), which virtualizes a physical device into multiple logical sensors, each with its own identity and radio budget.

Key technical components include:

  • MAC-Level Scheduling: They modified the Medium Access Control (MAC) layer in the Radio Access Network (RAN) to handle two new algorithms:
  • Downlink (DL): A credit-based weighted algorithm that spreads sensor data slots evenly over time to minimize jitter.
  • Uplink (UL): A deterministic slot ownership model where the base station (gNB) and device (UE) agree on a repeating cycle of "who owns which slot," eliminating the need for constant, resource-heavy signaling.
  • The SLED Framework: An end-to-end pipeline that carries these isolated streams through dedicated User Plane Functions (UPFs) to specific compute tiers, such as an Edge GPU, a Campus HPC, or a remote Leadership Computing Facility (LCF).
  • Agentic Coding Workflow: Perhaps most innovatively, the authors used AI agents (specifically Claude Opus) to write and debug the complex C code required for these RAN modifications. As shown in Fig. 3, the agent operated in a "modify-build-simulate" loop, drastically accelerating development.
Fig. 3: Agentic RAN development harness with the automated modify-build-simulate loop on HPC and human-orchestrated OTA validation on a live 5G/6G RAN testbed.

Datasets

The researchers evaluated SLED using a suite of seven realistic Sensor Workloads (detailed in Table III of the paper), including:

  • eMBB-VA: JPEG video analytics (15 Mbps).
  • URLLC-RT: Ultra-reliable low-latency robot telemetry (0.4 Mbps).
  • eMBB-BT: Bulk TCP transfers (50 Mbps).
  • eMBB-IQ: Raw radio signal streaming.

Evaluation was conducted on a physical Indoor OTA Testbed (Fig. 5) using USRP X300 software-defined radios and an RF simulator for scaling tests.

Key Insights and Ideas

  • The "Aha!" of Intra-UE Isolation: The paper proves that network-level slicing is insufficient if the bottleneck is the device's own internal scheduling. By moving the "slice" boundary inside the device's MAC layer, they achieved true end-to-end isolation.
  • AI as a Force Multiplier: The use of "agentic coding" wasn't just a gimmick. Once the AI successfully implemented SLED in one RAN stack (OpenAirInterface), it was able to port that logic to a second stack (OCUDU) in just a few hours. This suggests a future where complex protocol development is handled by AI agents overseen by human "orchestrators."
  • Deterministic Uplink is Essential: By using a pre-determined cycle for the uplink (Fig. 2), SLED avoids the "control overhead" that usually plagues 5G, making the system more efficient for high-speed sensor data.
Fig. 2: Sensor Slicing algorithm with credit-based DL slot allocation (top) and deterministic UL slot ownership (bottom).

Key Results

  • Latency Stability: In high-traffic scenarios, standard network slicing saw critical telemetry delays increase by over 2,300-fold. SLED, by contrast, maintained perfectly stable latency.
  • Development Speedup: Porting the feature to the OCUDU stack took only 2 hours and 40 minutes with the AI agent, compared to the 6.5 weeks it took for the initial manual/AI-assisted implementation in OAI—a 400x reduction in time.
  • Precision: Throughput for each sensor slice remained within 0.9% to 1.4% of the target allocation, even as the number of slices increased.
  • HPC Integration: The team successfully demonstrated a "Continual Learning" loop where video frames were sent to a remote HPC, processed, and returned with 99% delivery and stable round-trip times.

Limitations and Future Work

While SLED provides robust isolation, the authors note a few areas for growth:

  • Fixed Ratios: The current implementation uses pre-configured slice ratios. Future iterations could benefit from dynamic ratios that adapt to real-time sensor demand.
  • TDD Patterns: The experiments used a fixed Time Division Duplex (TDD) pattern; exploring how SLED interacts with more flexible or adaptive TDD patterns would be a valuable next step.
  • Scale: While they tested up to seven slices per device, real-world "fleets" of robots might require even higher density and more complex orchestration.

The SLED software and datasets are available at https://github.com/UNL-CPN-Lab/SLED.

Summary updated

Abstract

Modern sensing systems generate large volumes of heterogeneous data streams from multiple sensor modalities, increasing reliance on datacenters for computation, artificial intelligence/superintelligence (AI/SI) inference, and AI/SI training. Each sensor stream may have different network and compute requirements depending on its type and data volume. However, current wireless systems typically allocate resources at the device level, leading to contention among co-located sensor streams, where bulk data transfers may starve latency-critical telemetry. To this end, this paper introduces Sensor Slicing (SL). SL provides deterministic, medium access control (MAC)-level isolation for individual sensors within a single user equipment (UE) in open radio access networks (O-RAN). Using agentic coding techniques, SL is implemented in two O-RAN-compliant packages, reducing development time by 400 times (2 h 40 min [OCUDU] vs. 6.5 weeks [OAI]), 228-fold fewer human prompts (4 vs. 913), and less than half the tokens (0.64B vs. 1.43B). An associated end-to-end framework, SLED, is implemented to establish per-sensor radio and network isolation across edge, campus, and remote high-performance computing (HPC) tiers. Over-the-air wireless experiments demonstrate that while conventional network slicing causes critical delays to increase more than 2,300-fold under high traffic load, SLED maintains stable latency. Scaling evaluations on the two RAN stacks show that throughput remains within 0.9% and 1.4% of expected allocations even as the number of sensor slices increases. The results demonstrate that SLED preserves latency-critical telemetry while effectively utilizing resources across the sensor-to-HPC pipeline. The SLED software and evaluation datasets are publicly available at https://github.com/UNL-CPN-Lab/SLED-Sensor-SLicing-from-Edge-to-Datacenter.
BibTeX
@misc{subedi2026sled,
  author = {Alan Subedi and Avhishek Biswas and Apala Pramanik and Amlan Balabantaray and Mikhail Titov and Santosh Pitla and Ozgur Kilic and Mehmet C. Vuran},
  title = {{SLED: Sensor SLicing from Edge to Datacenter}},
  year = {2026},
  eprint = {2611.21937},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21937},
}
Figure