PAI-Field + Typed guardrails

In development

Measure the distance. Don't guess it.

PAI-Field measures how far every person is from the machine with 3D LiDAR, ten times a second, independently of the AI's own camera. Typed guardrails use that measurement to decide whether each AI command may run, and check again every 100 ms while the machine moves.

  1. 01

    3D LiDAR ×2–4

    Mounted high in the corners of the cell, looking down, so the machine never hides a person from all of them.

  2. 02

    PAI-Field

    Removes the empty cell and the machine itself from the point cloud, and measures what is left.

  3. 03

    Typed guardrails

    Eight checks on every command and on the running command every 100 ms: allow, slow, stop or reject.

  4. 04

    Edge Box

    Writes only what passed. Stops the outputs if guardrails go quiet.

Why LiDAR, and what it does not replace

2D camera (the AI's eyes)Safety laser scannerPAI-Field
DistanceEstimated from pixels. Shadows, glare and dust throw it offMeasuredMeasured (time of flight)
CoverageField of viewOne horizontal planeThe whole volume: crouching, reaching, climbing
Certified for safetyNoYes (IEC 61496, ISO 13849)No
JobPlan the workThe safety functionDecide which AI commands may run

Keep your safety laser scanner. PAI-Field sits on top of it. Point clouds carry no faces, so workers are not identified, which matters when you explain the system to the people on the floor.

What happens to every scan

Every 100 ms, for every sensor. Design target: under 50 ms from scan to distance.

  1. 01ReceiveEach LiDAR's points, timestamped. All sensors share one clock over PTP (IEEE 1588).
  2. 02AlignTransform into one cell coordinate frame, using the mounting positions found at commissioning.
  3. 03CropKeep only the monitored volume: the cell and the space around it.
  4. 04Subtract the empty cellCompare with the background recorded when the cell was empty. Keep only what was not there.
  5. 05Subtract the machineBuild the arm's current shape from its joint angles and remove those points.
  6. 06GroupCluster nearby points. Ignore clusters too small or too low to be a person.
  7. 07MeasureShortest distance from each cluster to the machine's hazard zone: the arm as it is now, plus where the current command is taking it.
  8. 08PublishSend the nearest distance, its timestamp and sensor health to Typed guardrails.

Anything unexpected counts as a person

PAI-Field does not need to recognize a person to stop for one. Any cluster that is not part of the empty cell or the machine is treated as a person: a worker, a cart, a box someone left behind. A classifier may label things for the dashboard, but it is never on the path that decides whether the machine stops. The same scan always gives the same distance.

What Typed guardrails receive

{
  "cell": "line-3",
  "captured_at_ms": 1760000000123,
  "nearest_person_m": 1.42,
  "objects": [{ "id": 7, "distance_m": 1.42, "height_m": 1.71, "points": 412 }],
  "health": { "sensors_ok": 3, "sensors_total": 3, "blocked": [], "drift_mm": 4 }
}

If any sensor is unhealthy, nearest_person_m is left out. No distance means stop. PAI-Field never assumes the remaining sensors can see everything.

When a sensor can't be trusted

Every case below withholds the distance, so the cell stops. A person checks the sensor before an operator reset.

What happensHow PAI-Field notices
Cable cut or power lostNo scan for 200 ms
Window dirty or coveredToo few returns from surfaces that should be visible in the background
Sensor knocked out of placeReference reflectors in the cell appear more than 5 cm from where they were
Clocks drift apartPTP sync lost, or a scan stamped in the future

Typed guardrails: the decision

Each AI command, and the running command every 100 ms, goes through eight checks in this order. The first one that fires decides. Same input, same decision. The demo shows each step with the value it saw.

  1. 01TypeKnown action, known fields, values in range. Anything else is dropped.
  2. 02Stop commandstop and stop_conveyor always pass.
  3. 03InterlockOnce stopped, nothing moves until an operator resets.
  4. 04ConfidenceBelow the threshold set for the cell, stop.
  5. 05Point cloudNo distance from PAI-Field, stop.
  6. 06FreshnessScan older than 200 ms, stop.
  7. 07Protective fieldPerson inside it, stop.
  8. 08Warning zonePerson inside it, cap the speed.

The allowed actions for a cell come from the Edge Box tag map, narrowed further to what the AI is allowed to do.

Sizing the zones

From a person moving to a stop command takes up to about 250 ms by design: one scan, processing, one guardrail cycle. That delay, plus the machine's own stopping time, goes into the size of the protective field, the same way ISO 13855 sizes separation distances. The 1.5 m and 3.0 m in our demo are for show. On a real cell we measure the stopping time with your safety engineer and size the zones from it.

What's in the field

Pilot systems use off-the-shelf parts. We will choose the exact models with our first design partners.

PartPlanWhy
LiDARLow-cost industrial 3D LiDAR, 360° horizontal, 10 Hz, IP67 class. Two to four per cellOne sensor always has a blind spot behind the machine
MountingCorners of the cell, 2.5–3.5 m high, looking downHeads and arms are seen before they reach the machine
ClockPTP across all sensors and the computerScans from several sensors are merged as one moment, and their age is measured
ComputerFanless industrial PC, same class as Edge Box. No GPU requiredNothing on the stop path needs machine learning
Reference reflectorsTwo or three, fixed inside the cellSo a moved sensor is noticed

Commissioning a cell

  1. 01Measure stopping timeWith your safety engineer. The zone sizes come from it.
  2. 02Mount and alignInstall the sensors and find their positions from the reference reflectors.
  3. 03Record the empty cellThe background PAI-Field subtracts from every scan.
  4. 04Define and approve zonesIn cell coordinates. Only a version your safety engineer approved goes live.
  5. 05Walk the cellSomeone walks every direction and height while we check a distance is reported throughout. The record is kept.

What PAI-Field does not do

  • Replace safety laser scanners, light curtains or any certified safety function.
  • Identify, track or rate individual workers. We leave it out on purpose, even where it would be possible.
  • Plan where the machine goes. PAI-Field only measures.

Build it with us

We are looking for teams putting AI-driven robots or conveyors next to people. Design partners shape the sensor layout and the guardrail policy for their cell.

Become a design partner