← → space · scroll · click dots
01 / 19 · Project Beacon
KAUST

Project Beacon · capabilities showcase

KAUST
× Lucid

People movement at scale · 3D scans of real environments · robot, conveyors, sensors on the scan.

Presenter Mohammed-iliès Ayachi mohammed.ayachi@kaust.edu.sa
Principal Investigator Prof. Shehab Ahmed shehab.ahmed@kaust.edu.sa
02 / 19 · Agenda
scope of demonstration

Programme overview.

01
People movement & digital-twin behaviour. GPU-accelerated crowd model running in the browser; reinforcement-learned destination policies over the KAUST campus twin (CitiesOS, COP28); 37 published behavioural rules under the hood. Live demonstration.
02
3D-Gaussian capture of industrial environments. ANPERC main lab and Angle Lab reconstructed from handheld scan; rendered in the browser. Live walk-through.
03
Simulation & scene editing on captured geometry. Robotic cell, conveyors and sensors composited over the 3D-Gaussian scan; direct subtraction and CAD insertion in the captured scene. Live demonstration.
04
Edge computer vision in production. BitXaminer drill-bit inspection (Aramco) · GenSet rig analytics and drillstring dynamics (Aramco) · LineEye factory inspection (HiRoll Technology).
05
Strategic robotics partner programme. Active collaboration with United Robotics Group (Germany) — autonomous mobile robots, lab-automation manipulators, and humanoid platform.
03 / 19 · People movement
what we built

A simulator for how people actually move through a space.

The established crowd-simulation tools in this space are mostly offline planning suites — desktop builds, fixed scenarios, batch runs. We took the same problem and built something that runs live, in the browser, on a real 3D scan of the space — interactive while the floor plan is still being argued about.

Where it could matter on a factory floor

  • Worker flow — finding bottlenecks, comparing line layouts.
  • Mixed traffic — workers and AGVs (automated guided vehicles — driverless tugs and pallet movers running along defined plant routes) sharing the same aisles.
  • Evacuation — egress simulation on the real plant geometry.

What's our particular bet

  • Real-time on the GPU — change a parameter and watch the crowd respond in the same second.
  • Runs in a browser — no install, no workstation, anyone can open it.
  • Sits on a 3D scan — agents walk through the photographed space, not an authored CAD model.
04 / 19 · Digital twin
cognitive agents · COP28

Crowd cognition at urban scale.

CitiesOS demonstrated at the COP28 Climate Action Innovation Zone

KAUST digital twin

Agents don't follow waypoints. Each one carries a destination policy — population segments commute to work, students go to class, services attract residents on a daily rhythm.

Buildings emit attraction signals that bias agent decisions in real time, parameterised by time-of-day, capacity, and function.

Policies are reinforcement-learning trained on real KAUST campus movement data.

COP28 · KAUST booth · CitiesOS live demo

05 / 19 · Live
demo
01

Live · People movement

A live crowd,
in the browser.

Show
The Tawaf scenario — mass flow around a central obstacle, tens of thousands of agents.
Then
The general simulator — agents reacting to doors, bottlenecks, obstacles. Sliders tweak crowd behaviour live.
06 / 19 · Under the hood
how it works

Thirty-seven rules. Every agent, every frame.

A · Routing & goal 3 rules
01Flow-field gradient descent
02Driving force toward goal
03Wave activation (staggered start)
C · Aisle congestion 3 rules
11Density slowdown (Fruin)
12Visibility-based fallback
31Stop-and-go coordination gate
E · Walls, fixtures & lanes 3 rules
17Soft wall / obstacle repulsion
18Position rejection (clearance)
19World boundary
F · Hazards, supervisors, mobile equipment 3 rules
20Attractor agent (muster point / supervisor)
21Repulsor agent (hazard zone)
22Mobile collision agent (e.g. forklift / AGV)
B · Worker / AMR traffic 9 rules
04Anisotropic repulsion
05Anticipatory braking
06Anticipatory steering
07Reynolds alignment
08Gap-seeking lateral move
09Asymmetry breaker (gated noise)
10Leader drafting
29Lane-formation bias (right side)
30Tangential friction at contact
D · Physical handoffs & contact 4 rules
13Contact separation impulse
14Pre-drive contact strip
15Post-drive contact strip
16Trampling slowdown
G · Physical limits (force · speed · inertia) 6 rules
23Force magnitude cap
24Inertia cap (Δv per frame)
25Absolute speed cap
26Adaptive damping
27Impatience accumulator
28Fall state machine (4 states)
H · Emergent line behaviour 6 rules
32Freezing-by-heating noise
33Mass-derived body radius
34Stop-event detector
35Crowd-pressure heatmap
36Social groups (V/U formation)
37Herding parameter (direction blend)
Tawaf · the 37 rules running on real-world geometry

Engine. All 37 run in parallel on the GPU, every frame, for every agent. A physics layer resolves separation and collision alongside. Everything runs in the browser — no install. Scales by swapping the geometry.

07 / 19 · Combined
crowd × 3D scan

Our crowd simulation, on a real 3D scan.

Discovery agents populating the captured bay
Flow same scene, with traffic and bottlenecks
08 / 19 · 3D scans of real spaces
capture, don't author

We capture reality, not authored CAD.

Real industrial spaces scanned with a handheld camera or drone, reconstructed into a photorealistic 3D environment, delivered as a browser experience. One pipeline, several fidelities depending on what the use case needs.

How we capture

  • Spherical 360° capture for fast walkthroughs.
  • Multi-view photogrammetry for high-detail surfaces.
  • LIDAR-fused capture for measurement-grade geometry.

Where we've shipped this

ANPERC high-bay NEOM Digital Twin CitiesOS SmartGrid typhoon

A captured ground truth of what is actually built — so a digital model can be compared against, and registered to, the real environment.

09 / 19 · Live
demo
02

Live · 3D scan walk-through

Walk a real bay,
in the browser.

Show
Walk mode through the ANPERC scan. Point out the photorealism — pipes, ladders, electrical panels: none of it is modelled, all photographed.
Note
Captured in an afternoon. Reconstructed on Shaheen III. In a browser the next morning.
— —
10 / 19 · Live
demo
03

Live · Industrial cell on the scan

Robot, conveyors,
sensors. Live.

Show
The scene loads, the cell instantiates on top. Run the cycle. A part crosses the light barrier; the sensor fires. The KPI dashboard updates. Click any drive or sensor to inspect it.
Then
Switch to walk mode and walk around the cell — the scan stays behind everything.
11 / 19 · Putting it together
setup for the last demo

Scan reality. Simulate operations. Deliver in a browser.

The next demo combines the previous two pieces on one screen — a virtual process walkthrough of an installed cell, on the actual scanned environment, before any equipment moves. Directly applicable to Beacon §3.4 — operator training, digital work-instruction validation, and pre-build line review.

  • The 3D scan — our ANPERC bay, captured environment, no CAD.
  • A simulated factory cell on top — collaborative robot, conveyors, light-barrier sensors, gripper, parts flowing through.
  • A control-logic layer — sequence containers, signal routing, recording and replay of motion sequences.
  • Operator KPI dashboard — overlay on the browser, no install required.
  • Walk mode — anyone in the room can navigate the cell as if walking the floor.
ANPERC highbay scan · simulated robot cell on top · click to open
12 / 19 · Scene editing
subtract · insert

The captured scene is editable.

Subtract and insert ANPERC bay · robotic-arm rig dropped into a cleared volume

Scene authoring

Photogrammetric captures are usually read-only — what you scanned is what you get.

Ours aren't. We can subtract any volume from the captured 3D scene, then insert authored CAD — a robotic arm, a conveyor, a fixture — into the resulting void.

The same browser viewer, on the same real environment, now hosting whatever needs to be prototyped against it before anyone moves a piece of equipment.

13 / 19 · BitXaminer
production · Aramco

From a worn drill-bit to a classified defect.

BitXaminer — drill-bit inspection deployed at Aramco. Multi-view photogrammetric capture reconstructs each bit at sub-millimetre fidelity — individual cutter elements on a ~10 mm cutter zone, micro-chips, and fracture lines are resolved and graded. Same reconstruction discipline maps directly to Beacon §3.1's 0.10 mm electrical-clip target.

Capture pipeline retrieval · cleaning · multi-view photogrammetric scan
Inspection app photoreal 3D reconstruction · zoom to broken cutter
14 / 19 · BitXaminer · next phase
automation roadmap

BitXaminer · full-station automation, every bit size.

The next phase replaces every manual handoff with a closed-loop facility. A conveyor + overhead gantry walks each drill-bit through loading → cleaning → coloring → scanning → unloading; a robotic arm and adaptive camera rig position the optics for any bit geometry — no human in the dimensional loop.

BitXaminer automated facility · concept renders

What changes Drill-bits flow through the line on a conveyor + overhead gantry. Loading and unloading stations sit at opposite ends; cleaning, coloring, and scanning stations bracket the middle.
Adaptive optics A robotic arm repositions the camera rig for each bit size and geometry. The scan capture protocol becomes invariant to bit class.
Why it matters Closes the loop from retrieval → classified defect record with no manual handoff. Throughput stops being limited by an operator's day.
15 / 19 · LineEye
production · partner

LineEye — computer vision in a box.

Inline inspection deployed with HiRoll Technology. Three detection methods running concurrently on a single NVIDIA Jetson edge unit — plug-and-play on the line, no programming required.

LineEye · datasheet & live inspection app

01

Barcode

Detect & decode the barcode. Orientation is derived from the result — if the barcode reads correctly the bottle is upright; if not, it's flipped.

02

Contour

Analyses the bottle's silhouette to determine which side the handle and cap face — at line speed, no fiducials required.

03

HSV

Top-view colour segmentation isolates the bottle body and cap. Robust under variable line lighting.

LineEye unit

  • NVIDIA Jetson · CPU + GPU, deep-learning inference on-device
  • Multi-camera · real-time 2D & 3D (depth) capture, compression & streaming
  • Rugged chassis · passive cooling, factory-floor temperature range
  • Plug-and-play setup · no programming, no extensive training
Read the LineEye brochure (PDF) ↗
16 / 19 · Edge CV at a rig
production · Aramco × KAUST

Edge computer vision, shipped to a working rig.

GenSet Analytics — a camera and Jetson box mounted on the rig deck, watching four diesel generators and reporting each one's load class every five seconds. The same edge-CV discipline AME needs for inline anomaly detection at AMP2.

ADES drilling rig · solar-powered camera trailer · PTZ cam + Jetson Xavier NX

Solar-powered camera trailer at the drilling rig site, with a callout zoom on the row of generator exhaust pipes and lids that the on-device detector watches.

What the camera watches · exhaust lid angle → load class

17 / 19 · Drillstring dynamics
edge CV · Aramco × KAUST

Drillstring dynamics, from a single camera.

Same edge-CV camera, different inference: video stabilisation → ROI tracking → signal generation → downhole RPM and vibration estimates. Stick-slip and twist-off events surface on a dashboard without putting any new sensor downhole.

01 Frame integrity capture + corruption check; auto-reconnect on stream loss
02 Purpose-trained detector single-stage real-time CNN, 1,600+ annotated frames across day / night / occlusion
03 Tracker & association custom logic pairs detected objects across frames; stable IDs
04 Occlusion logic flags blocked targets instead of guessing — no silent bad data
05 Signal → class / estimate per-target calibration; project-specific inference head
06 Status & heartbeat streamed to backend; operator dashboard refresh every 5 s

Edge inference

  • NVIDIA Jetson Xavier NX
  • TensorRT-compiled detector graph
  • PTZ camera · on-device buffering

Services & data

  • Python AI worker → Django REST
  • PostgreSQL (status · heartbeat · cameras)
  • React operator dashboard

Ops & deploy

  • Docker Compose — one-command spin-up
  • Auto-restart on unexpected shutdown
  • 3-day per-station calibration protocol
Failure-mode
discipline
AI worker down → backend marks DOWN after grace window.  Camera stream lost → reconnect loop, 5 attempts × 5 s, recover-on-success.  Backend down → dashboard banner + inference buffered locally; no data loss on the next sync.
01 · input
Pipe rotation, captured live
02 · process
On-device computer-vision pipeline: Frame t-1 / Frame t → video stabilisation → detector → ROI → signal generation (with hop length) → RPM and vibration datapoints → signal accumulation → RPM and vibration estimators.
Pipeline · frame → stabilise → ROI → signal → estimator
03 · output
Time-series plot of downhole RPM (max/mean/min) with top-drive RPM overlaid, and downhole torque vs top-drive torque below.
Downhole RPM & torque · derived from video alone
18 / 19 · United Robotics Group
strategic partner · DE

Robotics partner — United Robotics Group.

KAUST robotics collaboration with United Robotics Group (Germany) — a unified portfolio spanning autonomous mobile robots, lab-automation manipulators, and a humanoid programme. Direct pathway into Beacon §3.2 (humanoid & advanced robotics).

uLog Lift — an advanced lifting autonomous mobile robot from United Robotics Group, designed to automate and optimise human-handling tasks across industries.
uLog Lift advanced lifting AMR · industrial material handling
uMobileLab — a mobile lab-automation manipulator from United Robotics Group.
uMobileLab mobile lab-automation manipulator
United Robotics Group humanoid programme — featured in German manufacturing press.
Humanoid programme URG bipedal platform · advanced robotics roadmap
19 / 19 · Mapping to Project Beacon
capability → workstream

How this maps to Project Beacon.

§3.1 · Primary near-term

Vision Anomaly Detection

Sub-millimetre photogrammetric reconstruction (BitXaminer). Edge CV in production with calibration discipline (GenSet, Drillstring). Multi-method on-line inspection (LineEye / HiRoll). Architecture pattern transfers to AMP2's electrical-clip station at the 0.10 mm target.

→ slides 13 · 14 · 15 · 16 · 17

§3.2 · Robotics

Humanoid & Advanced Robotics

Active partnership with United Robotics Group (DE) — autonomous mobile robots, lab-automation manipulators, humanoid platform. Robotic cells composited directly into captured plant geometry for prototyping prior to physical install.

→ slides 12 · 18

§3.3 · Digital Twin

Digital Twin & Simulation

GPU pedestrian / AMR-traffic model with reinforcement-learned destination policies over the KAUST campus twin (CitiesOS). 3D-Gaussian capture of real industrial space, executed at scale on Shaheen III.

→ slides 03 · 04 · 06 · 07 · 08

§3.4 · AR / VR

AR / VR for Manufacturing

Browser-renderable 3D-Gaussian scenes used as virtual process walkthroughs — robotic cells, conveyors, sensors composited on the captured environment. Operator training and pre-build line review without bespoke hardware.

→ slides 10 · 11 · 12

Omniverse compatibility. 3D-Gaussian scenes we capture at KAUST are natively importable into NVIDIA Omniverse via the NuRec libraries (3DGS → OpenUSD), so they sit alongside Lucid's existing Omniverse work on the AMP1 Powertrain Pack Line rather than replacing it.

Questions ?

Mohammed-iliès Ayachi · mohammed.ayachi@kaust.edu.sa
Prof. Shehab Ahmed · shehab.ahmed@kaust.edu.sa