Augmented reality in telecommunication covers two different jobs. The first is operators using AR in their own operations: a customer-care agent guiding a subscriber through a router fix over live video, or a technician following step-by-step overlays at a site. The second is operators carrying AR traffic for their customers over 5G and edge networks, where AR is promoted as one of the main applications the network enables. The first job, in customer care, is deployed and has published operator results. The second is where pilots disappoint. An offloaded AR frame has to get from the camera to an edge server and back within one frame time, 16.7 ms at 60 frames per second. On a live radio network, handovers cause recurring delay spikes and 5G bandwidth fluctuates far more than 4G.
What does AR in telecom mean, and which use cases are live?
Augmented reality places virtual objects over the real world, so the user sees both together, usually through a phone screen or an optical see-through display in glasses. VR differs in how much of the view is digital: AR keeps real elements as the main focus and adds digital layers on top. In telecom, the sources describe three kinds of AR use case.
- Visual support in customer care: the customer shows the device and issue over live video and the agent guides them with on-screen AR annotations.
- Field maintenance and remote expert guidance: technicians wear headsets or glasses that show device information and repair steps.
- AR and XR as a network service: the operator carries latency-critical immersive traffic, which is what trials such as Verizon and Ericsson’s L4S work target.
Customer care: letting the agent see the device
Vodafone Portugal adopted Vodafone Visual Support, built on TechSee. During a call the customer gets a link, starts a live video session with no download, app or login, and the agent guides them using on-screen AR annotations. Vodafone Portugal reported a 4% rise in First Call Resolution, a 10-point gain in NPS and 12% fewer technician dispatches. TechSee’s case study for Vodafone Group describes the same aims: higher First Contact Resolution and a lower dispatch rate. We read these as figures announced by an operator about a vendor’s product, not as an independent evaluation.
How are telecom operators using AR for field-technician remote assistance and network-operations overlays?
Field AR makes the information and expert knowledge available at the repair site. Deutsche Telekom’s T-Systems “AR for Maintenance” fed device information and repair guidance to HoloLens-style glasses, showed the work step by step, indicated which tools to use and could reorder spare parts automatically. The same article called such solutions “still in their infancy” and predicted that 5G would make them real from 2020, so read it as a dated, named operator example rather than as evidence of current scale. AT&T has developed a field-service AR solution for industrial and enterprise settings, with headsets that use 5G. Softweb Solutions cites an Ericsson report claiming that AR and VR for remote assistance can cut maintenance time by up to 50% and improve first-time fix rates by up to 30%. That figure is quoted second-hand from an Ericsson report.
When overlays are anchored to equipment in 3D and must stay registered as the technician moves, the session needs continuous tracking from motion sensors and processing to keep virtual objects aligned with reality. Overlays for network operations centres are a different topic; camera-based tooling on the network side is covered in computer vision in modern telecommunications.
How does edge computing change the latency budget from sensor to display?
Edge computing moves heavy processing off the headset to more powerful network servers. Offloading makes sense because glass-type AR devices have only 0.5 to 2 W of power available, against 3 to 7 W for a VR or AR head-mounted display, and 3GPP’s TR 26.998 expects some processing to move to a tethered phone or a server. The end-to-end time for an offloaded frame is the uplink transfer plus inference on the edge server plus the downlink return, and each transfer includes half the round-trip time. Published edge AI latency requirements cover uplink and downlink communication plus edge computing, not network delay alone.
Where the milliseconds go
| Budget item | Figure | Source context |
|---|---|---|
| Motion-to-photon limit | AR under 10 ms, VR under 20 ms | Survey table of edge AI offloading requirements |
| Motion sickness threshold | Rendered image latency over 15 ms can cause it | Wireless VR analysis |
| Sensor delay | Under 1 ms | VR budget breakdown |
| Display delay | About 10 to 15 ms, expected to fall to 5 ms | Leaves about 14 ms for compute and communication |
| Offloaded frame, 60 / 30 FPS | 16.7 ms / 33.3 ms, capture to result | Edge-assisted AR object detection |
| 4G loopback, ideal conditions | 25 ms | |
| MEC communication delay | Expected under 1 ms in metropolitan areas | Projection, not a measurement |
| Measured 5G mmWave RTT, best case | 14.5 to 14.7 ms, static and facing the base station | Boston uplink study |
Worked example: one frame over the best mmWave link
Northeastern University’s 5G mmWave uplink study assumes a 450 KB camera frame and its best measured conditions: 158 Mbps uplink and a 14.5 ms RTT. The working: 450 KB is 3.6 Mb, which takes about 22.8 ms to send at 158 Mbps, and adding the 14.5 ms RTT gives roughly 37 ms. The study reaches the same figure, more than twice the 60 FPS budget, before any inference runs.
Compression narrows the gap but, on our arithmetic, does not close it. In the MASCOTS ‘23 study, H.264 at 12 Mbps reduces the frame to about 50 KB with less than 2% loss of accuracy, and encoding takes about 6.3 ms on a Samsung Note 20. On the Boston study’s best link, 50 KB is 0.4 Mb, or about 2.5 ms of transfer. Adding the 14.5 ms RTT and 6.3 ms of encoding gives about 23 ms, which still exceeds 16.7 ms before inference and rendering. Purdue and Charter’s EdgeSys ‘22 study found that the gain from LTE to 5G came mainly from the RTT falling from 30 ms to 12 ms, while the bandwidth increase cut only about 10 ms. With compressed frames, round-trip time matters more than bandwidth.
3GPP’s split-rendering model has the edge pre-render the viewport, and the device then warps the image using the latest pose. On the device side, a Rutgers system ran motion-vector tracking in 2.24 ms on a Jetson TX2 and left more than 14 ms to render within the 16.7 ms frame. It was tested only over Wi-Fi, though, and its authors did not study outdoor channels, jitter or dropped packets. TR 26.998 allows 50 ms pose-to-render-to-photon. We read that looser figure as possible because of the local warp, not as permission for a slow network loop. Placement also matters: Huawei’s Cloud VR white paper says the rendering server must sit on the network edge closest to the user.
How much uplink does an AR session need on a mobile network?
Camera-based AR sends a video stream in the uplink. Cameras on the glasses send video to an XR server, which estimates the user’s pose with SLAM. Because medium-quality video is enough for SLAM, the uplink stream is often downscaled compared with the downlink. Northeastern’s researchers argue that this pattern conflicts with 5G mmWave’s asymmetric, downlink-first design and needs uplink-centric operation.
Measured mmWave uplink degrades sharply outside ideal conditions. In Boston, 5G mmWave reached about three times LTE’s uplink bandwidth and latency when the phone faced the base station. Facing away, walking or driving, it fell to about LTE’s level or below. Static uplink facing the base station measured 131.8 to 157.7 Mbps. Walking away measured 50.5 Mbps, and driving measured 39.7 ± 21.2 Mbps with a 33.0 ± 22.0 ms RTT. Static LTE managed 53.4 Mbps. Average mmWave uplink was 47.21 Mbps in Chicago and 43.98 Mbps in Indianapolis, against about 150 Mbps in Boston.
The required figure depends heavily on the design. The MASCOTS ‘23 study estimates that naive edge-assisted object detection would need a stable 273 to 1331 Mbps of uplink, depending on the server GPU, which it considers unlikely to be available in the near future. Lighter perception tasks need far less uplink. In the TARMM study, object detection from a Meta Quest 3 sent about 22 KB uplink per frame at 20 FPS. That works out to about 440 KB/s, or roughly 3.5 Mbps. Semantic segmentation sent 110 KB per frame.
Not every XR workload is limited by the uplink. Nokia’s 3GPP-calibrated simulations found the modelled XR cases limited by the downlink, with only six devices per cell for 30 Mbit/s AR/VR in dense-urban FR1 under a 10 ms packet delay budget. We read the two findings as describing different workloads. Perception offload sends camera frames up. Rendered streaming sends video down. A pilot’s first design question is which of the two it is.
What bitrate does streaming VR or XR content need, and what does delivery cost the operator?
Streamed XR needs tens to hundreds of megabits per second per user, with frame-level timing. A 4K AR/VR stream needs more than 60 Mbps, packet delivery within 15 ms and frame loss below 1%. Wider estimates run from 100 Mbps for entry-level 1K VR to 1 Gbps at compressed human-eye resolution. Multi-user 360-degree 3D at 120 FPS and 4K per eye needs at least 100 Mbps downlink per user and under 20 ms end to end. Huawei places advanced 12K VR at 200 Mbps to 1 Gbps and notes a trade-off. Field-of-view streaming cuts bandwidth, but it moves network delay into the 20 ms interaction budget, whereas full-view streaming leaves that budget on the terminal.
Frames are bursty: the peak-to-average frame size is 2, so the radio must be planned for peaks. In the same multi-connectivity study, a single AR/VR stream used about 15% of an FR1 cell’s link resources. On volume, 60 Mbps for one hour is 216,000 Mb, or 27 GB, and 100 Mbps for an hour is 45 GB. In the wireless VR study, edge caching of pre-rendered frames for popular actions reduced computing delay.
What are 5G’s practical limitations for AR/VR, such as mmWave coverage, uplink capacity and congestion on the live RAN?
mmWave links are fragile. The human body attenuates mmWave signals by about 20 to 35 dB, and diffraction is minimal, so the best conditions need line of sight. In a multi-connectivity model, a single FR2 gNB missed the AR/VR frame-loss target beyond 130 m, and under blockage FR2 data rates can fall below FR1. Measured beamforming caused throughput drops in 25 to 30% of cases. Uplink throughput under lateral motion swung from 115 Mbps to 0, and the authors call these causes fundamental to mmWave, not fixable by tuning.
Throughput variance can limit AR even when the mean bandwidth is high. Real commercial traces show that 5G offers higher mean bandwidth than 3G or 4G with much larger fluctuations, which makes it hard for congestion control to adapt. BBR relies on 6 to 10 RTT bandwidth estimation windows and reacts too slowly to millisecond-scale channel changes. In the MASCOTS ‘23 driving scenario, average throughput was 64.02 Mbps, against an estimated requirement of 31.44 Mbps. The ±45.58 Mbps standard deviation still prevented satisfactory accuracy. So a link that cleared the requirement on average still failed the application in that study.
There is also a fair counterpoint to the “5G unlocks AR” story. Pre-5G app-level techniques, frame compression and local tracking, raise accuracy on both LTE and mmWave, and the extra gain from mmWave over LTE was marginal. The EdgeSys ‘22 study found that a simple design using H.264 and Lucas-Kanade tracking came within 3.07% of ideal offloading accuracy on LTE and 1.51% on 5G mmWave. For many telecom AR use cases, application engineering does more than the radio generation.
Where do telecom AR pilots fail: latency, throughput, content distribution or devices?
The table sets the four candidates against what a lab hides.
| Failure point | How it shows on a live network | Why the lab misses it |
|---|---|---|
| Latency tails from handover | In an indoor 5G O-RAN testbed, queueing delay below 20 ms in steady state rises above 200 ms at handover, even under light load | A static rig never changes cell |
| Handover interruption | 20 to 50 ms for basic and conditional handover, 2 to 10 ms with DAPS | Harmful when handovers are frequent |
| Throughput variance | Apps reduce frame rate or resolution, or freeze | Best case is a static phone facing the base station |
| Packet loss (fixed-line evidence) | Huawei reports 10⁻⁵ to 10⁻⁷ loss in the lab against 10⁻³ to 10⁻⁵ live, on copper lines | One late packet loses the whole frame |
| Device fragmentation | 3GPP defines standalone, edge-dependent and tethered AR device types; a relayed tethered loop crosses four wireless links | The pilot uses one device on one link |
| Content distribution | Caching pre-rendered frames for popular actions cut computing delay |
In the TARMM 5G O-RAN testbed, 99th-percentile RTT rose from 40 ms to 124 ms, and the delay spikes recurred at every handover. For segmentation offload, 95th-percentile latency went above 100 ms. Tethering adds its own problem, because the session’s QoS must include the glass-to-phone link for bitrate, loss, delay and jitter. Our own pattern write-up on why AR/VR pilots stall in production covers the hardware and content side. Live testing is the right moment to apply it.
Verizon and Ericsson reported up to 50% latency reduction from L4S signalling in the RAN. That trial ran in Ericsson’s Santa Clara lab on Verizon’s C-Band spectrum, not on a live network.
How to test a telecom AR pilot before rollout
The test plan should measure the conditions that break the budget, not the average that passes it.
- Write the budget end to end. Give capture, encode, uplink, inference, downlink, render and display a number each, at the target frame rate.
- Set a tail target, not a mean. The VR reliability work uses a probability bound: delay over 20 ms in no more than 1% of frames. 3GPP’s XR evaluations count a user as satisfied only when 99% of packets arrive within the delay budget.
- Walk across cell boundaries. Measure queueing delay and RTT through handover events, as the TARMM study did, not on a single cell.
- Turn away from the antenna and move. Repeat sessions facing away, walking laterally and driving.
- Test more than one site. The same mmWave measurement method gave about 150 Mbps in Boston and under 50 Mbps in Chicago and Indianapolis.
- Across six 5G traces, the Hera protocol showed up to 50% lower delay than BBR and 66% lower than Cubic in some of the traces.
- Measure the tethering link separately. TR 26.998 names the UE’s media session handler as best placed to test latency, loss and bitrate on that link.
- Join RAN counters to application QoE. Per-UE queueing delay, PRB usage and drop rate are available from the RAN, and erratic VR display is hard to diagnose from alarms alone.
Joining RAN counters to QoE matters commercially as well. Huawei notes that many dissatisfied VR users unsubscribe without complaining. The per-subscriber telemetry needed to catch that is the same material covered in big data in telecom.
The edge rendering node
Once the budget shows how many milliseconds the edge server gets, the server itself becomes a sizing question. Our companion pages cover the CPU side: multi-core vs single-core processors for edge AR/VR rendering. For the board itself, our guide to EPYC 4005 motherboard selection covers platform choice for edge rendering servers. The MASCOTS ‘23 study adds one point to keep in mind. A faster server GPU shortens inference and so relaxes the network requirement, but even with faster GPUs and app-level optimisations, neither LTE nor mmWave reached satisfactory accuracy in the driving scenario.
What the evidence says to build first
One telecom AR use with published operator outcomes is annotated video in customer care, at Vodafone Portugal. World-locked, offloaded AR over mmWave is where measured uplink, handover tails and blockage still work against the design. It should be engineered around those limits, with local tracking, compression and on-device warping, rather than assuming 5G will cover the gap. An open question remains whether DAPS handover and L4S cut latency on live networks as their trials and specifications report.
FAQ
Is augmented reality a type of AI?
AR is a way of displaying virtual objects composited with the real world. Many augmented reality applications depend on AI, though: object detection with deep neural networks is a core building block and the most computationally demanding one.
Does edge computing really reduce the delay of waiting for a cloud server?
Yes, by shortening the distance to computing resources. MEC is expected to cut communication delay to under 1 ms in metropolitan areas. Measured AR gains from LTE to 5G came mostly from a shorter RTT, not from more bandwidth.
What is the difference between mmWave 5G and regular 5G?
mmWave uses much higher frequencies with wide channels, so it can be very fast with line of sight but is easily blocked, even by a human body. The lower FR1 band gives wider coverage and higher reliability, while FR2 gives high rates only when conditions are good.