VisionFlow Robotics

Submission — Edge Innovation Hub 2026

Autonomous visual intelligence for hazardous water-treatment environments

VisionFlow Robotics deploys small, modular autonomous robots with edge-AI computer vision to detect chemical leaks, cathodic protection failures, and filter clogging — before they become incidents.

See the live demoRead the technical brief

01 / The problem

Three structural gaps in current water-treatment monitoring

Static sensors have blind spots

Fixed instrumentation captures process data but cannot observe the physical state of equipment. Corrosion, biofouling and structural drift develop in zones no transducer monitors.

Human inspection means human exposure

Confined spaces, chlorine atmospheres and energised electrolytic cells create routine risk for technicians. Inspection cadence is constrained by safety procedure, not by asset need.

Early-stage faults stay invisible until they cascade

By the time a process variable shifts, the underlying defect — a hairline crack, a pinhole leak, a fouled membrane — has already progressed. Reactive maintenance dominates.

02 / The solution

An autonomous fleet that sees what fixed sensors cannot

A coordinated fleet of compact rovers patrols defined inspection routes through treatment halls, electrolyser corridors and filtration galleries. Each unit carries a multi-modal sensor stack and runs computer-vision models on-board.

Edge inference removes the requirement for continuous video uplink. Only structured detections — fault class, location, severity, timestamp — are forwarded to the plant control layer over OPC-UA or MQTT.

The system is hardware-agnostic on the robot side and integrates with existing SCADA, CMMS and historian infrastructure without replacement.

VisionFlow architecture
Layer 01
Robot fleet
▶
Layer 02
Edge inference
▶
Layer 03
Plant SCADA
Modality
RGB + thermal + LiDAR
Latency
< 80 ms on-device
Bandwidth
Metadata only uplink

Deployment Architecture

VisionFlow Robotics — Deployment ArchitectureIndustrial autonomous inspection system · Edge AI · No cloud dependencyLAYER 03 — PLANT CONTROL SYSTEMSSCADAPlant control layerReal-time process monitoringAlert integrationCMMSMaintenance managementAuto work-order generationMaintenance schedulingOperator dashboardSeverity heatmapHistorical trend analysisMulti-site overviewStructured alerts onlyOPC-UA · MQTT · REST · metadata only, no videoLAYER 02 — EDGE COMPUTING · ON-BOARD JETSON ORIN NANOAll inference runs locally · <80ms latency · No internet requiredCorrosion detectionYOLOv8m-seg modelSeverity score 0–100Affected area %Pixel-level segmentationClassification: Minimal→CriticalChemical leak detectionHSV color-space analysisNo ML model requiredWet surface anomalyChemical discolorationStatus: Seepage→Critical leakFilter clog detectionTexture + brightness analysisClog score 0–100Flow reduction estimate %Recommendation auto-generatedStatus: Clean→Replace immediatelyCamera feed · on-device onlyCamera feed · on-device onlyLAYER 01 — ROBOT FLEET · AUTONOMOUS PATROLModular design · 4–6h battery · SLAM navigation · Hazardous environment ratedFixed camera + edge boxPoC Phase 1 — quick winMount on existing structure3–5 critical inspection pointsJetson Orin Nano in weatherproof boxReady week 1 of PoCNo robot hardware requiredMobile autonomous robotPoC Phase 2 — primary targetPatrols predefined routes autonomouslyRGB + thermal + LiDAR sensor stackSLAM navigation · obstacle avoidance4h patrol · modular designPoC validation targetDrone platformPost-PoC roadmap · Year 2Elevated structure inspectionConfined space accessTank interior inspectionSame edge AI stackSame software · new platformPlant systemsEdge AI modelsPoC primary targetAlternative platformsVisionFlow Robotics · Edge 2026
  • Layer 1: Robot fleet patrols inspection routes — fixed cameras for PoC phase 1, mobile robots for phase 2.
  • Layer 2: All AI inference runs on-board the robot — no video ever leaves the device.
  • Layer 3: Only structured alerts sent via OPC-UA or MQTT to existing plant SCADA and CMMS systems.

03 / Live demo

Three inspection modes, one autonomous platform

Live demo — all 5 inspection modules active

visionflow.demo / corrosionRGB · pixel-wise mask
Open in full screen →

Upload a metallic-surface image; the model returns a corrosion mask and a severity estimate. Same inference pipeline runs unmodified on the on-board edge accelerator.

04 / Why us

Built for the harshest industrial environments

Industry challenge
VisionFlow solution
Human exposure to hazardous zones
+Autonomous patrol removes humans from risk
Static sensors with blind spots
+Continuous visual coverage, every angle
Reactive, scheduled maintenance
+Predictive alerts before failures cascade
Siloed inspection data
+Structured data direct to SCADA/CMMS

05 / Technical specifications

Two deployment modes — same AI core

Recommended PoC start

Phase 1 — Edge box deployment

PoC months 1–3 · Ready week one

  • Connects to existing industrial cameras
  • Weatherproof Jetson Orin Nano enclosure
  • No new infrastructure required
  • Immediate SCADA integration
  • Cost: ~$1,500 per inspection point
Full autonomy target

Phase 2 — Mobile robot

PoC months 4–6 · Autonomous patrol

  • Compact wheeled platform · 480mm · 12kg
  • RGB + Thermal + LiDAR sensor stack
  • SLAM navigation · obstacle avoidance
  • IP65 rated · 4h battery
  • Same software stack as Phase 1
Phase 2 robot reference design — top viewscale 1 : 8
RGB CAMERA · 4K · 90° FOVTHERMAL · 320×240 · 8–14μmLiDAR · 360° · 16-CHEDGE COMPUTE · ORIN NANOIP65 · 4 h BATTERY · 12 kg480 mm
Edge hardware
Phase 1 · Edge box

Jetson Orin Nano · weatherproof box

Wall-mounted enclosure wired to existing CCTV / inspection cameras. 15 W envelope, PoE or 24 V DC, industrial temperature grade.

Phase 2 · Mobile robot

Jetson Orin Nano + Hailo-8 · on-board

Same compute stack carried on the robot. Up to 40 TOPS on-device inference, battery-powered, dual-redundant on critical units.

Models
Phase 1 · Edge box

Same AI core · fixed-camera feeds

YOLOv8 segmentation and the custom corrosion classifier run unchanged against scheduled stills from fixed cameras.

Phase 2 · Mobile robot

Same AI core · mobile RGB + thermal

Identical models, fed by the robot's RGB, thermal and LiDAR stack during autonomous patrols. 120k-image proprietary training set.

Protocols
Phase 1 · Edge box

Wired OPC-UA / MQTT to SCADA

Ethernet into plant LAN, direct outbound to Siemens, Rockwell or Schneider PLC stacks. TLS, no ingress required.

Phase 2 · Mobile robot

Wi-Fi 6 / 5G mesh → same SCADA

Robot streams structured alerts over Wi-Fi 6 or private 5G to an edge gateway, then OPC-UA / MQTT to the same SCADA endpoints.

Autonomy
Phase 1 · Edge box

Fixed install · scheduled triggers

No motion. Inference runs on cron or event triggers from existing cameras — zero navigation required to go live.

Phase 2 · Mobile robot

SLAM navigation · 4 h battery

Visual-inertial localisation, obstacle avoidance, auto-docking and inductive recharge between patrol cycles.

06 / Traction & roadmap

A staged path from validated prototype to multi-site operation

  1. 01
    Lab

    Prototype validated

  2. 02
    Months 1–6

    Edge PoC deployment

  3. 03
    Months 7–12

    Pilot deployment

  4. 04
    Year 2

    Multi-site rollout

07 / Team

Image placeholder

Portrait

Chief Executive
Name placeholder
Image placeholder

Portrait

Chief Technology
Name placeholder
Image placeholder

Portrait

Head of Robotics
Name placeholder
Image placeholder

Portrait

Head of Computer Vision
Name placeholder

08 / Contact

Direct
contact@visionflow-robotics.com

For technical evaluation, partnership discussion, or to schedule a controlled demonstration on representative imagery, please use the form or write directly.