Most bin picking projects do not stall on the robot. They do not stall on the gripper, and they rarely stall on the sensor. They stall because nobody can produce a usable 3D model of the part.

If you have scoped a vision-guided picking cell, you have probably hit this. The application is sound, the ROI works, the robot is selected, and then someone asks for the CAD file of the part and the project goes quiet for three weeks. This article covers why the dependency exists, what "unusable model" actually means in practice, and the four documented paths to a working pick when the file is missing.

Why bin picking wants a CAD model in the first place

The dependency is not arbitrary. Photoneo publishes the Bin Picking Studio setup as an eight-step workflow, and step three is uploading the object. Photoneo's own description of that step is direct: all you need to start planning your bin picking gripping points is a CAD model of the product you want to pick.

That model does more work than most people assume. It carries four separate jobs:

• Localization. Photoneo's localization engine is CAD-based. It matches the scanned point cloud against the model geometry to work out where each instance of the part is and how it is oriented.
• Gripping point definition. You select where and how the gripper grasps, directly on the model, in a virtual environment. No model, no grip points.
• Collision checking. Path planning needs the part's geometry to know whether the gripper can reach a given instance without hitting its neighbors or the bin wall.
• Simulation and debugging. The virtual environment that lets you validate a solution before committing hardware is built from the same geometry.

Remove the model and you have not lost one feature. You have lost the setup workflow.

Why the file is missing, and why "we have the CAD" often is not true

In our experience quoting these cells, the model problem shows up in five distinct forms, and only the first is the one people expect.

Situation	Why it blocks
No file exists	Purchased casting, forging, or legacy component. The supplier never provided a model and may not have one.
The supplier will not release it	The geometry is treated as proprietary. Common with contract manufacturing and with customer-supplied parts.
The model is as-designed, not as-manufactured	Draft angles, flash, warp, and weld distortion mean the real part does not match nominal geometry closely enough for reliable matching.
The model is too heavy	A full assembly with modeled threads, internal features, and fastener detail carries geometry the localizer will never see from outside the part.
The part deforms	Bags, textiles, gaskets, and soft goods have no stable geometry for a rigid model to match against.

The middle three are the ones that catch teams out, because everyone answers "yes, we have the CAD" and then discovers three weeks in that having a file and having a usable model are different conditions.

Path one: scan the part and build the model from the scan

If a physical sample exists, you do not need the original file. You need a scan good enough to generate geometry from.

Photoneo sells this path directly. Their positioning for Photoneo 3D Meshing opens with "No CAD? No problem," and describes creating a model of any object and then transforming it into a workable CAD model ready for your projects. The input is a scan. The output is geometry the localization engine can use.

What makes or breaks this path is the quality of the underlying scan, and that is a sensor decision rather than a software one. Three properties matter:

• Point density. Determines how much real geometry survives into the mesh. On the PhoXi 3D Scanner Gen3, point-to-point distance at the sweet spot runs from 0.16 mm on the S up to 0.83 mm on the XL. A feature smaller than your point spacing does not exist in the model.
• Data completeness on your material. A scan with dropouts produces a mesh with holes, and holes in the wrong places remove exactly the features the localizer would have matched on.
• Noise floor. Temporal noise at the Gen3 sweet spot ranges from 0.03 mm on the S to 0.21 mm on the XL. Noise becomes surface roughness in the mesh, and surface roughness becomes false matches.

Path two: build the model inside the software

This one is new, and for some applications it collapses the whole problem into a single afternoon.

Bin Picking Studio 1.12, released in July 2026, lets you create the model natively. You connect the sensor, trigger an acquisition, and generate the model inside the software, then refine it with integrated mesh editing tools that crop out noise, remove secondary objects, apply basic filtering, and adjust face normals and the origin point. Photoneo added .PLY import alongside existing formats in the same release.

The limit is worth reading carefully, because it decides whether this path applies to you at all. Photoneo describes in-app modeling as ideal for structured or semi-structured applications where objects are consistently viewed from the same angle the model was created from. For fully random object orientations, they state they still recommend creating the 3D model using Photoneo 3D Meshing.

Randomly oriented parts in a bin are precisely the case that recommendation is about. So if your application is trays, blisters, racks, or fixtured parts, this is now the fastest route to a working model. If it is a genuinely random bin, Path one remains the one Photoneo points to.

Path three: skip the model entirely with AI localization

For some applications the answer is not a better model but no model. Photoneo's AI-driven piece picking is positioned the same way, opening with "No CAD? No problem," and stating that shape and size are not the constraint.

This is the right path when geometry is genuinely unstable or genuinely unbounded: mixed-SKU fulfillment, deformable goods, and cases where you are picking hundreds of part numbers rather than four. Photoneo's AnyPick module for boxes and the LayerPick delayering module extend this into logistics.

The trade is specificity. CAD-based localization gives you a precise pose for a known part, which is what you need for assembly, machine loading, or anything that has to be placed to a tolerance. AI localization gives you a pickable object, which is what you need when the population is open-ended.

Path four: fix the model you already have

If a file exists but localization is slow or unreliable, the model itself is often the problem. The localizer only ever sees the outside of the part from one viewpoint at a time, so anything the scanner cannot observe is dead weight in the matching process.

Practical corrections, in rough order of impact: strip internal geometry the sensor will never see, remove modeled threads and fastener detail, delete assembly context so the model represents one part rather than a subassembly, and reduce polygon count to something proportionate to your point density rather than to your CAD kernel's default tessellation.

As of Bin Picking Studio 1.12 you can do a good deal of this without leaving the software. The same integrated mesh editing tools that support Path two also apply to imported models: crop, remove secondary objects, filter, and adjust face normals and the origin point. Photoneo also changed the renderer so that back-faces of CAD meshes now display in black rather than being invisible, which gives immediate visual feedback on normal orientation. Anyone who has spent a day debugging a model that localized inside-out will recognize why that matters.

There is also a case for scanning a part you already have a file for. If the as-manufactured part differs meaningfully from nominal because of draft, flash, warp, or weld distortion, a mesh built from a real sample will match your real parts better than the designer's intent does.

Choosing between the four

Your situation	Start here
Parts sit in trays, blisters, racks, or fixtures, and are consistently oriented.	In-app model creation. Fastest route, and Photoneo recommends it for exactly this case.
Small, stable set of rigid parts in a random bin. No file, but you have samples.	Scan and mesh with Photoneo 3D Meshing, then CAD-based localization.
File exists but localization is slow or drops parts.	Simplify the model first, using the integrated mesh editing tools. Rescan a real part if as-built differs from nominal.
Hundreds of SKUs, or the population changes weekly.	AI localization. Modeling each part will not scale.
Parts deform, compress, or drape.	AI localization. Rigid geometry matching has nothing stable to match.
Placement has to hit a tolerance downstream.	CAD-based localization, by whichever route gets you the model. You need pose, not just a pickable object.

What solving the model problem does not solve

The sections above draw on Photoneo's published documentation. What follows is Automation Distribution's assessment, based on specifying and supporting these systems.

A model gets you to a localization result. It does not get you to a working cell, and three things downstream of it fail often enough to be worth budgeting for.

• Occlusion. A perfect model does not help if the part you want is buried, or if the feature that defines its orientation faces away from the sensor. Photoneo's MultiView module addresses this by merging scans from multiple viewpoints, and specifically names thin, C-shaped, and L-shaped sheet metal, features obscured in a single view, and self-occluding objects among the cases it is built for.
• Calibration. The pose your localizer computes is only as good as the transform between sensor and robot. Budget for a marker board or calibration ball at specification time rather than discovering the need during commissioning.
• Reachability. Localizing a part your robot cannot reach at that orientation, or cannot grip without collision, produces a cell that finds parts and picks none of them. This is a path planning and gripper question, not a vision question.

Frequently asked questions

Do you need a CAD model for bin picking?

For CAD-based localization, yes. Photoneo's Bin Picking Studio workflow requires uploading a model of the object, which the system uses for localization, gripping point definition, collision checking, and simulation. But a CAD file is not the only way to get that model. You can generate one by scanning a physical sample, create one inside Bin Picking Studio as of version 1.12, or bypass the requirement entirely using AI localization, which Photoneo offers for parts where shape and size are not fixed.

Can I create the 3D model inside Bin Picking Studio?

Yes, as of Bin Picking Studio 1.12. You connect the sensor, trigger an acquisition, and generate the model natively, then refine it with integrated mesh editing tools. Photoneo describes this as ideal for structured or semi-structured applications where objects are consistently viewed from the same angle the model was created from. For fully random object orientations, Photoneo states they still recommend creating the model using Photoneo 3D Meshing software.

Can you create a 3D model of a part by scanning it?

Yes. Photoneo 3D Meshing takes scan data and produces a model, which Photoneo states can then be transformed into a workable CAD model. The practical limit is scan quality: features finer than your point spacing will not appear, and dropouts on difficult surfaces become holes in the mesh.

Why is my CAD model making localization slow?

Usually because it carries geometry the sensor cannot observe. Modeled threads, internal features, fastener detail, and full assembly context all add matching cost without adding matchable surface. Reducing the model to the external geometry of a single part, at a polygon density proportionate to your scan resolution, is generally the highest-impact fix. As of version 1.12, the integrated mesh editing tools let you do much of this without leaving the software.

Should I scan a part even if I have the CAD file?

Sometimes. If the as-manufactured part diverges from nominal geometry because of draft angles, flash, warp, or weld distortion, a mesh built from a real sample represents what the sensor will actually see. Castings, forgings, and weldments are the usual candidates.

Does the laser wavelength matter for scanning parts to model them?

On difficult materials it is the dominant variable. Photoneo states blue laser technology delivers up to 2.5 times better 3D data completeness on complex materials, naming black plastics, reflective metals, rubber, matte finishes, and translucent surfaces. Across the Gen3 datasheets, blue and red variants publish identical range, accuracy, planarity, noise, and point density figures, so the choice is about how the projected pattern behaves on your surface rather than about raw performance.

What if my parts do not have stable geometry at all?

Then rigid model matching is the wrong tool. Bags, textiles, gaskets, and other deformable goods change shape between scans, so there is nothing consistent for a model to match against. AI localization is the documented path for those applications.