1 · Perceive
A YOLOv5 model on the Rockchip RV1126B NPU (RKNN) detects the person, per frame, probabilistic and noisy — 0.85–0.93 confidence on real footage.
Solution · Smart Machine Safety Bundle
AI perception in. Machine safety out.
The Smart Machine Safety Bundle turns AI perception — person proximity from cameras, anomalies from sound and vibration — into deterministic, PLC-ready safety decisions for industrial machine cells: gate, slow or stop, alert, and record. Below is the whole loop running on real footage, with the actual EdgeEmbed Runtime deciding — not a scripted demo. Evidence-ready, never certified — the certification is the customer's.
What's happening
Perceive. Decide. Act.
Three stages, one deterministic engine in the middle. The perception side is probabilistic and can be swapped for any model; the decide step is always the same EdgeEmbed Runtime.
2 · Decide
EdgeEmbed Runtime evaluates the machine-safety policy against the detection — the same deterministic engine behind every EdgeEmbed product, not a demo shortcut.
3 · Act
One policy-admitted action reaches the machine — slow, stop, alarm, or resume — and every decision is recorded for replay.
Implementation
What's under the hood
Every technology in the loop, from lens to actuator — and the line in the middle where probability stops and determinism starts.
Probabilistic — perceptionDeterministic — decision & control
Real factory footage. Frames in, no intelligence yet.
videoOpenCV
Stock person detector — never retrained for safety.
YOLOv5sCOCOONNX
Quantized and compiled for the edge SoC.
RV1126Brknn-toolkit2int8
Boxes become confidence-carrying safety events.
zone mapIoU trackinghysteresisfail-safe
Post-inference event normalization
Evaluates the bundle policy. Never sees a pixel.
Model Gatepolicy.jsonrecorderC11
One admitted action — latched until operator reset.
STOPSLOWalarmlatch
The real decision log
Every dispatch, straight from the engine
This is genuine output from EdgeEmbed Runtime running the machine_safety_bundle policy against the detections above — not a mock-up.
runtime_runner — machine_safety_bundle
EdgeEmbed Runtime: 2.2.0 (real engine)
EVENT t=0.9s VISION.PERSON_IN_WARNING_ZONE confidence=0.88
-> DISPATCH machine_safety::MACHINE_SLOW
-> DISPATCH machine_safety::SHOW_YELLOW_WARNING
-> DISPATCH machine_safety::PLAY_WARNING_BEEP
EVENT t=5.6s VISION.PERSON_IN_DANGER_ZONE confidence=0.85
-> DISPATCH machine_safety::MACHINE_STOP
-> DISPATCH machine_safety::SHOW_RED_ALERT
-> DISPATCH machine_safety::PLAY_CRITICAL_ALARM
-> DISPATCH machine_safety::REQUIRE_OPERATOR_RESET
EVENT t=13.0s DEMO.OPERATOR_RESET
-> DISPATCH machine_safety::MACHINE_RUN
-> DISPATCH machine_safety::SHOW_GREEN_STATUS
Final: events_submitted=5 completed=17 failed=0
Condensed from a full run against real footage — the engine dispatched 17 actions across 5 events with zero failures.
What's in the bundle
Policy, connection, and a pilot in a box
Everything a machine cell needs between the AI perception you choose and the controllers you already run.
Authorable site policy
A base machine-safety policy pack plus a per-site overlay: proximity zones, anomaly thresholds, and a degraded-mode ladder are authored per site and rendered into the concrete bundle the machine runs.
Industrial connect packs
Decisions leave as control-ready signals: CAN, Modbus/PLC — validated against an independent third-party PLC stack with a real ladder-logic scan cycle — and MQTT alerts northbound to your monitoring.
Pilot in a box
Provision a site policy, install runtime + bundle as a system service, and run a smoke scenario through the real engine — producing a verifiable evidence record. Validated on real industrial ARM hardware.
The reference design
This is how every bundle on the runtime is built
The bundle proves the platform pattern: a new domain is a policy pack plus thin connect plugins — no changes to the deterministic core. Study it once, then build your own.
Behavior as data
Everything the cell is allowed to do lives in the policy pack — reviewable by a safety engineer, versioned like code, authored per site.
Connect, never replace
The bundle drives the installed base — PLCs, CAN devices, monitoring — through thin plugins. The safety chain you trust stays exactly where it is.
The same deterministic core
Under the bundle runs the same engine as every EdgeEmbed product: gated decisions, bounded latency, and a recorded, replayable trail for every action.
Pilots
Bringing AI safety to a machine cell?
We're selecting a small number of design partners for machine-safety pilots. If you have a cell, a camera, and a PLC, we would love to talk.