TORUS is an Unattended Facility Sensing platform for hyperscale, colocation, and enterprise AI/HPC data centers. It monitors campus perimeters, data halls, and critical facility approaches across ground, visual, and drone/UAS detection layers, classifies security events at the edge with on-device AI, and drives facility response actuation under safety interlock, even where mains power and network access are unavailable.
The platform is designed to support TIA-942 and OCP-aligned facility programs, and PacketFive Design Services can adapt TORUS for other customer requirements.
AI/HPC data center facilities are exposed at campus approaches, service yards, utility corridors, and data hall boundaries long before a door alarm or fence contact. Those areas are exactly where power, network, and staffing are constrained. Fixed cameras see a cone; patrols see a moment; fences report only contact.
TORUS holds those approaches continuously and passively. It detects intrusion indicators in more than one physical domain, correlates the cues, and hands facility operators a classified security event with location and confidence, not a raw alarm.
Every security event is captured, sequenced, and delivered. Coverage gaps are detected, not discovered after the fact.
Sensing emits nothing to find. The only active element is a response node, and it actuates only through an interlock.
Battery and solar operation, long-range radio, months between visits.
Every detection and every control action is timestamped and logged for post-incident review.
TORUS covers an approach from below, at eye level, and overhead. Each domain matures on its own track, so the ground layer deploys today and the system extends without replacing what is already emplaced.
Seismic and acoustic detection on unattended battery nodes reporting over long-range radio. Pre-production hardware complete.
Long-range thermal cores and LiDAR on a mast, cued by ground detections to identify and track at standoff, day or night.
Radar-class RF detection of drones and UAS activity over the facility. An active research track: long-range air detection demands higher transmit power and dedicated RF front ends.
A deployment mixes node classes to match cost, event type, and facility constraints. All classes report on the same event contract, so the operations layer treats a mixed field as one sensor network.
| # | Class | Link / RF | Maturity | Role & trade-off |
|---|---|---|---|---|
| 1 | Fixed Wireless | LoRa 433 MHz | FIELDED | Lowest cost per node; the default perimeter element for wide, dense rings. |
| 2 | Programmable Wireless | SDR (AD9371 / AD9363) + LoRa failsafe |
R&D | Software-defined RF for experimental, R&D, and high-availability facility use. Ships with LoRa as a failsafe that can also serve as the main link; where LoRa alone suffices, class 1 is the cheaper choice. |
| 3 | Wired Optical | Optical fibre | R&D | Emanation-quiet, tamper-resistant wired transport for the highest-assurance segments. |
| 4 | Transducer | per host node | CONCEPT | Response node. On a validated cue it actuates a facility response through a mandatory safety interlock (activate, safety-validate, then actuate). |
Every sensing channel on the TORUS-SN ground node is a named, datasheet-backed component selected for low standby power and field serviceability.
4.5 Hz vertical geophone into a low-noise front end with 128-step digital gain and a hardware wake comparator that trips with the processor asleep. Signatures: footfall, vehicles, digging.
Two 24-bit I2S microphones give a stereo field for on-device classification. Audio is classified at the node; raw audio never leaves it.
LoRa 433 MHz uplink, GNSS scanning, and passive Wi-Fi-scan positioning from one device. A relocated node reports its own new position.
Dual-core edge MCU with on-device ML, temperature-compensated timekeeping, single-cell Li-ion with USB-C charging, and three independently gated sensing rails.
The runtime guarantee is a reliability property, not a slogan: every event is captured, ordered, and delivered, and the highest-severity event is never queued behind noise.
Each node stamps every event with a monotonic sequence number.
Events are forwarded with de-duplication on identity; nothing is silently dropped.
A break in the sequence is flagged and replayed, so a missed event is detected, not discovered later.
Critical security events preempt lower-priority processing in the event queue.
On complexity, stated correctly. The runtime is a hard-real-time, zero-miss pipeline. The genuinely NP-hard problem in TORUS is optimal sensor placement for complete coverage (a set-cover / art-gallery problem). That is why the deployment is rehearsed in simulation first: the coverage problem is approximated off-line, then zero-miss processing is guaranteed on-line.
A representative emplacement: ground nodes form the outer detection ring, visual masts cover the likely approaches, the experimental air layer watches the volume overhead, and every element reports to a single gateway at the asset.
Data center campuses are expensive to survey twice. A deployment is rehearsed as a cyber-physical twin in NVIDIA Isaac Sim with terrain reconstructed from mapping data, the full system placed virtually, and unauthorized entry scenarios run against it. The figure below is real output from the open-source planner.
In campus deployments the twin earns its keep twice: the facility noise floor is simulated first, so seismic thresholds and acoustic classifiers arrive on site pre-tuned against the local background instead of by trial and error.
Mixed node classes sense passively and classify at the edge, reporting compact events over LoRa, SDR, or optical.
Receives the field, fuses and timestamps events on a Jetson-class edge computer, and forwards over flexible backhaul through the northbound integration API. Cluster-aware gateways chain and federate across distances for high availability. Off-grid, rugged.
The facility operations layer: one live common operating picture for the security desk, facilities team, and field staff, with transducer response under safety interlock.
Command, Control, Infrastructure, Security, Response, and Tier. The security desk, facilities team, and field operator share a single fused picture; each sees it in the form the job needs, and transducer nodes actuate only through the safety interlock.
Map wall with live node, gateway, and mast state; fused, prioritised alerts; triage and escalation; thermal/video cued from the masts; timeline replay; deployment planning with the Isaac rehearsal link; integrations for PSIM, DCIM, BMS, VMS, and SIEM.
Handheld map with own position and the node ring; prioritised alerts with bearing and distance; cue a sensor or mast; confirm or reclassify in the field; offline-tolerant sync over the gateway.
A transducer walks arm, safety-validate, then actuate. Actuation is refused until validation passes.
Who armed, validated, or actuated what, and when, is recorded for post-incident review.
The operations layer runs on-premises, with a fully air-gapped option for high-security facilities.
Ground sensing node, in detail.
Open →The four node classes and their trade-offs.
Open →Aggregation, fusion, backhaul.
Open →The facility operations layer.
Open →Intelligence at the edge, gateway, and operations.
Open →What is built, and the roadmap.
Open →Open-source hardware, firmware, design data.
GitHub ↗Deployment simulation.
GitHub ↗TORUS is an independent platform. Company names and product model numbers referenced across this brief and its datasheets (including but not limited to Teledyne FLIR, Semtech (LR1110), R.T. Clark, InvenSense (INMP441), Texas Instruments, Microchip, Espressif (ESP32-S3), Analog Devices (DS3231, AD9371, AD9363), and NVIDIA Jetson and Isaac Sim) are used solely for engineering and bill-of-materials identification. Their use does not imply any affiliation with, sponsorship by, or endorsement by those companies. Supply of any such third-party product or component to the designers, manufacturers, integrators, or evaluators of the TORUS platform remains at the sole discretion of the respective owning company, organisation, or legal entity. "FOR EVALUATION" is a document-handling marking only and does not denote any government classification. All trademarks are the property of their respective owners.