Skip to content
collimo Submit the case
A yellow robotic arm inside an enclosed cell, shown moving a panel along a line.

Machine vision software for robotic cells

We tell the robot where to grip. The client stays yours.

And what to reject on the line. We train the model on the actual parts from your customer.

Submit the case

Archival footage, not our system.

01 Your customer

We don’t interfere with your client.

Whoever sends us a project also hands over the relationship with their company. That’s worth more than the project. We treat it this way.

02 The boundary

Where your work ends and ours begins.

We write the inspection that deploys the model to production, on the robot and the PLC you already have. No hand-written rules: models that work when the part shifts position, or when the lighting changes in the shop.

You

  • Mechanics and EOAT.
  • Electrical panel and wiring.
  • PLC and safety.
  • Layout, assembly, on-site commissioning.
  • The relationship with the client.

Us

  • Image acquisition and dataset construction.
  • Model training and validation.
  • On-machine inference.
  • The interface to your PLC.
  • Model monitoring over time.

We communicate with the PLC via TCP/IP, Profinet, or EtherNet/IP. We deliver the model in ONNX. It runs on an industrial PC with a GPU inside the panel, not in the cloud: your clients’ images don’t leave the facility.

03 Before you promise it

Before promising it to the client, let us see it.

Send a photo of the part, the tolerance, and the cycle time. Within two working days, you’ll receive a two-page document with one of three answers. It’s free, and if the case doesn’t hold up, we’ll tell you, even if it costs us the job.

camera vision · robot guidance example on a sample part of ours
What the robot camera sees: a bin of metal brackets in bulk, viewed from above. The software has marked the detected parts with a bounding box and the selected pick with a green crosshair, where the gripper is descending.
candidates 7 out of 14 cycle 840 ms pick
This is what the robot-wrist camera sees on our test bin, not on a customer’s production line: the model outlines the parts it detects, ignores the occluded ones, and provides the gripper with the pick point. Drag to move the camera.

  1. Feasible.

    It states the threshold we guarantee and how we measure it during commissioning.

  2. Feasible with reservations.

    It specifies what needs to change first: lighting, grip, how the part arrives, cycle time.

  3. Not feasible.

    It explains why. If there’s another way to solve that problem, we’ll write it down even if we don’t handle it.

04 What stays with you

If you stop calling us tomorrow, the system keeps running.

  • We deliver the trained model exported. It runs without us.
  • The annotated datasets are yours, images included. We don’t reuse them for other jobs without asking.
  • The integration code is in your repository. Not ours.
  • The fee covers monitoring and retraining. If you cancel it, the system doesn’t stop: it just stops improving.
The same part broken down into four layers suspended one above the other, connected by vertical laser lines.

The reason the fee exists is that models degrade. The sheet metal supplier changes, the workshop lighting shifts in February, and the false reject rate rises without anyone touching anything. That’s what we watch.

05 Who is behind it

We are new, and we say it ourselves.

Collimo started within Firmamento and is currently a project, not a separate company. When it has enough work, it will stand on its own. We mention it here because it’s the kind of thing you’d find out anyway, and it’s better to read it from us.

We don’t have clients to show yet. That’s why the first step is free and the final payment is tied to a written threshold: we take the risk off your hands since we can’t yet do it with references.

06 The case

Submit the case.

Not much is needed. A photo of the part, even taken with a phone. What the robot needs to do. The tolerance and cycle time. If you have the drawing, better, but if not, we start anyway.

Two lines are enough. Painted sheet metal, bulk picking from a bin, scratch inspection.

Photo of the part or the drawing. Even a phone photo works. Up to three files, 20 MB total.

    seconds

    Why now

    We’re taking the first three projects under conditions we won’t repeat, because we need references to show. It’s not a countdown promotion: once the three are taken, it’s over.

    Example completed on our sample

    Feasibility verdict

    Part
    Milled aluminum bracket, curved surface, two through-holes
    Request
    Detect scratches and dents on the curved surface, part arriving in bulk
    Tolerance
    Defects from 0.15 mm depth and up
    Cycle
    6 seconds per part
    Outcome
    Feasible with reservations

    What works

    The curved surface isn’t a problem. With grazing light from two sides, a 0.15 mm scratch produces enough contrast, and the model distinguishes it from regular machining lines, which are uniform and oriented.

    What needs to change first

    What we guarantee

    With the three corrections above: at least 97% of defects over 0.15 mm detected, with no more than 3% false rejects, measured on 500 parts during commissioning, half of which are known defective. If we don’t meet this during commissioning, the final payment isn’t due.

    What’s not covered

    Defects under 0.15 mm, geometric deformations, and defects on the inner edges of the holes. Those would require different optics and another round of measurements.

    This is an example completed on our own part, to show what you receive. It’s not a client’s work.