Fusing Multi-Source Sensing Data into Latency-Compliant AI Decisions
Real-world RAN inference receives heterogeneous data from RF channels, ISAC radar, UE telemetry, environmental sensors, and network KPIs — simultaneously, asynchronously, in incompatible formats. InfraSense fuses them into a single coherent input, enforces a latency class across the entire pipeline — fusion included — and explicitly refuses any request it cannot honor within the declared bound. No silent degradation. No misrouting.
Patent-filed latency class architecture — enforced tier placement, explicit refusal
The OranSense Dashboard
Real-time xApp telemetry, RIC policy control, and AI-driven anomaly detection across your distributed RAN estate — all in one pane of glass.
0
0
3 updating6.8 bps/Hz
4.2 ms
| Name | Version | Status | CPU | Latency |
|---|---|---|---|---|
| QoS Optimizer | v2.4.1 | running | 12% | 1.8 ms |
| Interference Mitigator | v1.9.0 | running | 8% | 2.1 ms |
| Load Balancer | v3.1.2 | updating | — | — |
| Anomaly Detector | v1.2.5 | running | 21% | 3.4 ms |
| Energy Saver | v2.0.0 | standby | 0% | — |
High PRB utilization on n41 — sector C4-NE
2 min ago
xApp Load Balancer update scheduled — 02:00 UTC
14 min ago
Interference event cleared — gNB-0882
41 min ago
Ready to deploy OranSense? Bring AI-driven O-RAN intelligence to your network in days, not months.
View live dashboardOne governed pipeline. Fusion to decision. Latency enforced end-to-end.
InfraSense is a working demonstrator of a dedicated AI inference layer that treats the full pipeline — multi-source fusion through model inference to auditable control action — as a single latency-accountable unit. Every request declares a latency class. The orchestrator places it only on infrastructure tiers capable of honoring that class across the entire pipeline, fusion included. If no compliant placement exists, the request is explicitly refused rather than silently misrouted to a congested shared endpoint where it will degrade unpredictably. The architecture runs today on NVIDIA GPU infrastructure, attaches to existing O-RAN interfaces (A1, E2, E3, O1, O2) without replacing the stack, and requires no additional hardware.
- Multi-source fusion: RF channel measurements, ISAC radar returns, UE telemetry, environmental sensors, and network KPIs fused into one coherent inference input — with timestamp alignment, schema normalisation, and conflict resolution built in
- End-to-end latency enforcement: fusion time counts against the declared latency budget, not just model inference time — three enforced classes (RT <10ms, Near-RT 10ms–1s, Non-RT >1s) with hard tier placement rules
- Explicit refusal semantics: when no infrastructure tier can honor the declared latency class, the request is refused with a structured error — not silently degraded, not retried into a pass, not misrouted to a slower tier
- A1 SensingIntent policies: operators declare sensing objectives (target, accuracy, resource envelope) via a new A1 policy type; rApps and xApps derive configuration automatically and tune toward targets in a closed loop
- SMO Global Agent arbitration: when aggregate sensing commitments exceed network capacity, the agent resolves conflicts by priority, reports explicit infeasibility rather than silently over-committing, and re-plans when capacity changes
- Verifiable honesty: measured values are labelled measured, simulated values are labelled simulated — every benchmark call is a real model call whose failures surface as explicit errors, with runs traced on independent third-party infrastructure
NVIDIA Aerial — RF Signal Intelligence
GPU-accelerated 5G L1/L2 baseband + NIM inference
Three Innovations. One Governed Inference Layer.
InfraSense combines three components that run today, end-to-end, on NVIDIA GPU infrastructure — attaching to existing O-RAN interfaces without replacing the stack.
Dedicated Inference Layer
Receives requests comprising sensing data and a latency objective. Classifies each into a latency class (RT / Near-RT / Non-RT), selects an inference entity from a catalog subject to a hard tier rule — a latency class admits only infrastructure tiers capable of honoring it — and emits an auditable control action over the interface appropriate to its kind and tier (E2 / O1 / O2). Explicit refusal when no compliant placement exists.
A1 SensingIntent Policies
A new A1 policy type that lets operators declare sensing objectives: target entity, sensing modality, accuracy threshold, update rate, and resource envelope. An rApp translates each policy into xApp configuration. A closed-loop convergence mechanism monitors accuracy against the declared threshold and tunes configuration automatically — without operator intervention.
SMO Global Agent
Maintains a global view of all active SensingIntent policies and their aggregate resource demands. When commitments exceed capacity — the demonstrator exercises a 170% over-commitment scenario — the agent resolves conflicts by declared priority, downgrades lower-priority intents to a degraded-but-honest state, and reports explicit infeasibility rather than silently over-committing the network.
The core shift: from a passive, human-managed infrastructure to an active, autonomous NVIDIA-accelerated agent — enabling proactive, real-time optimisation rather than reactive management.
Industry Use Cases
Latency-guaranteed, fusion-aware AI inference transforms operations across every sector where the RAN is the sensing and decision layer.
Remote Health Monitoring
Continuous patient vitals, fall detection, and anomaly alerting over 5G RAN — where inference latency directly determines clinical response time.
Disaster Response & Mesh
Resilient edge AI networks that maintain sensing and inference capability when core infrastructure is degraded — prioritising life-safety control actions under constrained capacity.
Retail & Venue Intelligence
Real-time footfall analytics, queue detection, and personalised engagement — fusing RF sensing with camera and environmental data into sub-second inference decisions.
Smart Manufacturing
Closed-loop AI control for Industry 4.0 — fusing machine telemetry, environmental sensors, and network KPIs into latency-classified decisions that drive robotic and process control.
Energy & Smart Grid
AI-driven grid monitoring and fault prediction over private 5G — where sensing data from distributed assets must be fused and acted on within strict latency windows to prevent cascade failures.
Defense & Tactical Edge
Mission-critical sensing and inference at the tactical edge — where explicit refusal semantics and auditable control actions are as important as raw latency performance.
Real GPU Execution. Not Simulation.
InfraSense latency evidence is measured as application-level round-trips of real model calls — on two independent dedicated local serving stacks, with failures surfacing as explicit errors. No retries to mask failures. No simulated latency curves.
Narrow, Predictable Latency Band on Dedicated Local Inference
Dedicated local GPU inference — tested on vLLM slices on A100 MIG hardware and, after a hardware refresh, on a separate serving engine on an H100 — stayed within a narrow, predictable band across every run. The 1.3–1.8× max-to-min variance is the behavior a latency-guaranteed inference layer exists to deliver: a bound that can be declared, enforced, and refused against. This is what makes a latency class meaningful — not a target, but a guarantee.
Outcome
Dedicated local GPU serving reproduced narrow, predictable latency bands on two independent stacks — precisely the behavior a latency-guaranteed inference layer exists to deliver. The variance is tight enough to set a meaningful latency class bound.
1.3–1.8×
Max-to-min latency variance (dedicated GPU)
2
Independent serving engines validated (A100 MIG + H100)
Measured
Application-level round-trips of real model calls
Order-of-Magnitude Swings on Shared Cloud Under Congestion
Public shared-cloud model endpoints in InfraSense tests swung from approximately 1 second to tens of seconds on identical requests under normal conditions. During a measured congestion window in July 2026, individual calls stretched to 50–70 seconds — with some requests rejected outright. This is not an edge case. It is the normal failure mode of shared infrastructure under load, and it is precisely what InfraSense's placement rules exist to prevent.
Outcome
The failure mode — silent degradation on congested shared endpoints — is reproducible on demand. InfraSense's tier placement rules prevent it by refusing misrouted requests rather than letting them degrade unpredictably. A Near-RT request that cannot be honored in 10ms–1s is refused, not silently served at 50 seconds.
50–70s
Measured cloud latency under congestion (July 2026)
10×+
Run-to-run latency swing on identical requests
Explicit
Rejections surfaced as errors — not retried into a pass
13 Scenario Tests Covering Every Agent in the Demonstrator Catalog
The benchmark suite covers every agent in the demonstrator's catalog — 13 scenario tests as of July 2026 — spanning both single-shot inference subjects and genuine multi-step agentic model subjects that exercise the full pipeline from sensing input through fusion to control action. Runs are additionally traced on independent third-party infrastructure for verifiability. The design principle is verifiable honesty: measured values are labelled measured, simulated values are labelled simulated.
Outcome
Every benchmark call is a real model call. Failures — timeouts, rejections, wrong answers — surface as explicit errors. Nothing is retried into a pass. The suite is designed to be reproduced by any evaluator with access to the same infrastructure.
13
Benchmark scenarios covering the full demonstrator catalog (July 2026)
3rd-party
Independent trace verification on every run
Zero
Retries to mask failures — timeouts and rejections surface as explicit errors
Standards & Ecosystem Alignment
The Gaps InfraSense Addresses
Two problems the O-RAN ecosystem currently has no standard answer for — stated precisely.
An xApp developer who needs a 10ms inference answer has no portable way to declare that requirement and no guarantee the platform will honor it. The failure mode is silent: a model call that lands on a congested shared endpoint simply arrives late, and the control loop degrades unpredictably. There is no refusal, no error, no signal — just a missed deadline that the orchestrator never sees.
Inference Latency Gap
InfraSense — Problem Statement
ISAC has no operator-grade management story: there is no standard way to tell the network what to sense, at what accuracy, at what resource cost — or to arbitrate when sensing goals collide with communication goals. Without a management interface, ISAC cannot be offered as a service, cannot be priced, and cannot be governed. It remains a research capability with no operational path to deployment.
ISAC Management Gap
InfraSense — Problem Statement
InfraSense supplies the governance layer that dynamic compute sharing creates the need for: per-request latency classes, enforced tier placement, explicit refusal, and auditable control actions — plus the intent-driven sensing management dimension that no compute-sharing work addresses. It is not a model serving framework. It is the policy and enforcement layer that sits above model serving and below the xApp.
Positioning
InfraSense — Technical Architecture
Ready to explore InfraSense?
Request a technical briefing or live demonstrator walkthrough with our engineering team. We can walk through the benchmark data, the A1 SensingIntent policy model, and the SMO arbitration scenario in detail.
Get in TouchStay ahead of the 5G sensing curve
Operator briefings, ISAC research, and AI-RAN analysis — delivered to your inbox. No noise, no marketing. Just signal.