Can vision locate this part? A feasibility checklist before you teach

You are standing at the cell with a sample part in your hand and someone asks: can a camera find this part? Guessing wrong either loses the application or buys hardware that never converges. This checklist turns that judgement into five checks with numbers attached: how many pixels the part must span, what finish and background it needs, which features survive rotation, and how steady the lighting must stay. A worked example sizes a shiny aluminum part on a Universal Robots UR3, and a ten-minute bench test confirms the answer before anyone orders hardware.

What makes a part locatable

Vision locate is template matching, not recognition. The software compares the taught reference against the snapshot and returns a position and an angle that drive the pick — the mechanics are the same in every vision-guided pick and place cycle. So the question is never whether the camera knows the part. It is whether the snapshot contains features that are high-contrast, identical from part to part, and unique under rotation. Five properties decide it:

1. The part spans enough pixels in the field of view.
2. The finish gives stable gray values, not moving highlights.
3. The part separates from its background.
4. The feature set pins the part's angle uniquely.
5. The lighting geometry stays fixed between teach and run.

Fail one and the locate turns intermittent — the most expensive failure mode in a robot cell, because it passes at the demonstration and drifts in production. The next three sections put a number on each point.

Checklist stage 1 — the part on the bench

• Matte or controlled finish. Specular highlights move when the part or the light moves, and an oxidized lot reflects differently from a polished one. Shiny metal is workable: machined edges, drilled holes, parting lines and steps hold their gray values under a directed light. Never teach a highlight itself, and if the whole face is mirror-flat, plan for diffuse or polarized light from the start.
• Features that survive lot-to-lot variation. Cast texture, etch density and print wear all drift between batches. Teach geometry first and print second, and re-examine the template when the surface process or the label supplier changes.
• Separation from the real background. Lay the part on what it will actually sit on, not on a clean bench mat. The silhouette needs a clear gray step: a matte fixture in the opposite tone gets you there — black under bright metal, light gray under dark plastics and rubber.
• Rotation-unique features. A round flange with an even bolt circle returns an ambiguous angle, and the matcher will choose among mirror-image scores. An off-center hole, a notch, a keyway or an asymmetric print block pins the rotation.
• Teachable features on the top face. The camera sees one face. A 51 mm tall part with its distinguishing details on the side wall will not locate from above; either the features move to the top, or you locate the silhouette and take the angle from a top-face detail.

Checklist stage 2 — the numbers that settle size

Two numbers settle the size question before any camera is chosen.

Field of view first: the snapshot must cover the part plus the placement scatter on both sides, with a border of context around the part.

A loosely presented 127 mm part with ±10 mm scatter needs at least 147 mm of FOV; choosing 160 mm leaves about 6 mm per side of spare margin.

Pixels on part next: millimetres per pixel equal the FOV divided by the horizontal pixel count, and every other size number follows from that scale.

The planning thresholds: 100 pixels across or more returns a stable position and angle. Between 30 and 100 pixels the locate still works but the angle output turns noisy — 30 pixels is ten times the 3 pixel feature floor, the point where a template has almost nothing left to interpolate. The smallest feature you rely on needs 3-4 pixels, because three is the minimum for the software to interpolate an edge to subpixel precision.

For pure locating, a mono sensor is the simpler read: the MV-SC5060XM-08M-WBN is a 6 MP mono smart camera with an 8 mm lens on GigE. Take the snapshot with the arm stationary and its rolling shutter is no disadvantage — the checklist already requires a still arm.

Checklist stage 3 — the cell around the part

• A fixed lighting angle. The matcher compares gray patterns; move the light and the pattern moves with it, and the score falls. Mount the light to the camera or the fixture so teach and run share one geometry — lighting geometry is what sets image contrast.
• Ambient light under control. A window, a work lamp or a nearby stack light adds a gradient that moves through the day. Shield the snapshot area or time the exposure between sources; ambient light variation in pick and place has known fixes.
• A stationary snapshot. Take the image when the arm is still. Rolling shutter sensors distort a moving scene, and even global shutter sensors blur if the exposure runs long.
• Light, lens, exposure, in that order. When a locate scores well on the bench and poorly in the cell, the difference is almost always illumination; the stable order — lighting first, then lens, then exposure is the fastest route back to a good image.
• One safe teach pose. Snapshot height inside the robot's reach, focus set once, no hands in the workspace during tests.

When the finish fights you, an integrated light removes the largest variable. The LS-SMC-0063 carries an integrated polarized white light with a 16 mm lens and runs 60 fps over GigE, while the MV-SC6050C-08M-XBN pairs a 5 MP color sensor with a multi-color light source, so contrast can be tuned by wavelength on colored parts.

Worked example: a 127 mm shiny aluminum part on a UR3

The case, from a real application set: a shiny aluminum part 5 in (127 mm) across, presented loosely in the workspace, taught to a Universal Robots UR3 whose snapshot pose sits close to the UR3's maximum height. The finish is the worry, because oxidation varies from lot to lot.

The inputs. Part width 127 mm, placement scatter assumed at ±10 mm for a loosely presented part, and machined edges assumed at 0.5 mm wide or more. The camera is the MV-SC3050XC-16M-ATN-H, a 5 MP color vision sensor with a 1/1.45" sensor format and an M12 mount. Assume a 4:3 aspect ratio at the nominal 5 MP: √(5,000,000 × 4/3) ≈ 2,582 pixels across the wide side.

The rule in words: the FOV must cover the part plus scatter on both sides; from that FOV, millimetres per pixel give the pixels on part and the size of the smallest usable feature.

The 0.5 mm machined edges land at 0.5 / 0.062 ≈ 8 pixels, comfortably over the 3 pixel floor.

Result: about 2,050 pixels on part against the 100 pixel threshold, a 20× margin. Decision: size cannot fail this part. The checks that decide it are finish and background — teach the machined edges and holes, put the part on a matte black fixture, and let diffuse light kill the highlights. The number that moves the result is the FOV. Raise the snapshot pose to clear the UR3's height limit and the FOV widens with it: at 320 mm the scale doubles to 0.124 mm/px and the part still spans about 1,025 pixels, still a clear pass. The part only falls to the 100 pixel line near a 3,280 mm FOV, which no wrist camera reaches.

When the usual answer is wrong

• The printed feature tempts you. Print carries the highest contrast on the part, so it becomes the template by default. Print also wears, re-registers between label lots and smears with oil. Locate on geometry, keep print as a secondary region, and if print is the only feature, check three lots before committing.
• The finish varies more than the template expects. A polished lot and an oxidized lot can score far apart on the same template. The 5 in aluminum case above teaches on edges precisely because shine is the variable: diffuse the light and re-teach on geometry instead of enlarging the template.
• The part is long and thin. One part in the same application set was too long and too thin to teach as a whole; the workable answer was to teach a section of it. That is the standard method, not a workaround — see teaching long thin parts to a vision system by feature.
• The part is wider than the snapshot FOV. A 305 × 178 mm part does not fit a 160 mm FOV. Move to a fixed overhead camera sized to the part, or locate on a sub-feature and treat the rest as dead weight; matching a part larger than the field of view is its own technique.
• The part is near-symmetric. Mirror-image scores make the matcher pick an angle at random. Add an asymmetric mark to the part, or locate position only and let a compliant gripper absorb the rotation.

The ten-minute bench test

1. Place the part on the production background at the worst realistic pose — furthest offset, strongest tilt.
2. Light it exactly as the cell will, with room lights present or shielded as in production. A result taken under different light says nothing about the cell.
3. Take one snapshot and measure three values: pixels on part (100 or more), smallest taught feature (4 pixels or more), and silhouette gray separation (40-50 levels or more on an 8-bit image, so a threshold keeps margin over sensor noise).
4. Rotate the part through twelve steps of 30° and locate at each. The angle output should track the rotation, and position should drift by a small fraction of the pick tolerance.
5. Repeat with the oldest, most oxidized part and the newest. A flat score across the set means the template sits on geometry; a spreading score means re-teach and fix the light.

Ten minutes at a bench replaces weeks of tuning on the floor, and each failure names the checklist point to fix.

The decision for each case

Next step

Run the bench test first; it costs nothing and settles five arguments at once. Then size the hardware from your own numbers: the camera finder filters smart cameras by resolution and shutter type, and the configurator below builds the rest of the chain around the camera you pick. If a check fails and the part still has to work, send the bench snapshots through the contact page with the five answers, and scoping a machine vision system before you call an integrator carries the rest of the station decisions — camera, lens, light and controller as one choice.