3D bin picking for flat parts: choosing a vision approach

Random bin picking demos well on tangled castings and brackets — and then your bin turns out to hold flat parts. Stamped blanks, shims, plates and washers stack face to face, so the camera's job stops being finding a part and becomes proving that two identical parts lying on each other are two parts. That is a height measurement problem, and height resolution is exactly where 3D sensor technologies differ most. This guide commits to numbers: how much Z repeatability a given thickness demands, the field of view and standoff a real bin needs, structured light versus stereo versus laser galvanometry for thin stacks, and the point at which a feeder is the honest answer. For bulky or curved parts the rules are kinder; see bin picking non-planar parts.

Why flat parts are the hard case

A stack of flat parts gives the sensor exactly two cues. One is the height step where a blank overlaps its neighbour, and that step equals one part thickness: 2 mm parts produce 2 mm steps, 1 mm shims produce 1 mm steps. The other is edge detail — a corner radius, a chamfer, a burr line — measured in tenths of a millimetre. Identical parts touching along a straight edge show no intensity edge at all, so height is the only trustworthy separator.

Then the mechanical failures arrive. Two oiled blanks stick and travel as one. A picked part lands face-down where the process needs a face-up blank. Polished stock burns holes in the point cloud precisely along the edges you need. Curved parts forgive because their surfaces generate slope changes everywhere; a flat part offers one weak step and nothing else.

The decision up front

Commit by thickness and stack behaviour:

• 10 mm and thicker, or poured so parts rarely overlap: an active stereo 3D camera, or a 2D smart camera over a single-layer tray, is enough. Height steps are large and edges stay visible.
• 2–10 mm, overlapping stacks: projected structured light with Z repeatability at most one fifth of thickness, plus a layer-picking strategy. This is the volume case, and it is solvable.
• 1–2 mm, or polished metal at any thickness: laser galvanometer stereo. Plan a trial pour with your actual parts before you commit the cell.
• Under 1 mm: do not random-bin-pick. Layers sit inside the noise band of any sensor you would mount over an open bin; feed or magazine the parts.

Choose the gripper in the same decision: bellows vacuum for most blanks, a magnet pitched for single-part retention on steel, Bernoulli for thin sheet that must stay unmarked. For the physics behind each method, 3D scanning method selection compares them in inspection terms.

The numbers that drive the sensor choice

The division by five is noise margin: a step measured with noise equal to the step itself is invisible, and edge localisation needs the step to stand several noise bands above the surface. The division by three is sampling: an edge cue smaller than three pixels cannot be placed to a pixel, so a 1 mm chamfer asks for about 0.3 mm per pixel. The FOV margin covers bin placement tolerance and the robot's approach envelope over the rim. Z range must see from the empty bin floor to a full stack plus the tilt of a part leaning on the wall.

Worked example: 2 mm blanks in a 600 × 400 mm bin

Stampings 80 × 40 × 2 mm, tumbled into a bin with a 600 × 400 mm interior and 200 mm walls.

• Field of view: 600 + 2 × 75 = 750 mm across, 400 + 2 × 75 = 550 mm deep.
• Standoff: 1.0 m from sensor to bin floor; projected structured-light cameras usually work between 0.8 and 1.5 m.
• XY resolution: segmentation runs on height, so XY only has to place pose and clearance. At 0.5 mm per pixel a 40 mm blank spans 80 pixels, and the FOV needs 750 / 0.5 = 1500 by 550 / 0.5 = 1100 pixels — 1.65 MP, so buy a 2 MP sensor.
• Z: 2 / 5 = 0.4 mm repeatability is the ceiling; a 0.2 mm class sensor halves the risk on oiled or tilted stacks.
• Z range: 200 wall + 180 stack + 50 margin = 430 mm; specify 450 mm or more.
• Cycle: an 8-pattern capture at 25 fps takes 0.32 s, cloud build and match about 1 s, and you plan five picks per scan; with a 5 s robot move that is roughly 6 s per part, about 600 parts per hour from one robot.

Run your own bin through the calculator:

Structured light, stereo or laser galvanometer

These bands are the published classes for each technology; the data-sheet figure of the sensor you buy is the number you hold it to at acceptance.

In our range, the projected structured-light route is the MV-DPS200P-02 and the laser-galvanometer route is the MV-DLS300P-04.

Laser galvanometer and profiler sensors are lasers classified under IEC 60825: mount the aperture above eye level or behind the cell guarding, and keep the interlock in the circuit.

Laser profilers earn their keep downstream, where blanks arrive fixtured on a conveyor or tray and the part moves under the sensor:

A profiler rastering a whole random bin needs thousands of profiles and a moving axis, so it loses the cycle race to an area sensor. For a tray that already presents parts in rows, its micron-class height noise is decisive, and laser line extraction noise control is what turns that height into a clean edge.

Shiny flat stock: getting a point cloud at all

Polished or oiled blanks act like mirrors: projected light reflects away from the camera and the cloud grows holes along exactly the edges that separate parts. Four fixes, in order of preference:

1. Laser galvanometer stereo — the scanning line with per-line high-dynamic-range exposure survives specular flashes that blank a full-frame pattern.
2. Blue structured light with two or three bracketed exposures per capture; the shorter wavelength scatters less on machined steel.
3. Cross-polarization, at the cost of roughly half your returned light — budget exposure or gain headroom for it.
4. Move the sensor off the mirror angle: 15–25° of tilt between projector axis and surface normal sends the glare away from the lens.

Matte spray on a trial pour proves the software works, but it is a lab tool, not a production answer. For metallic parts specifically, the 3D bin picking systems for metallic parts overview collects the sensor-side options used on bare metal; the same glare physics drives stereo depth on dark stock, where the projector, not the part, supplies the contrast.

Layer picking, flip stations and the single-part check

Three strategies turn a marginal flat-part cell into a working one.

Layer picking. Accept only parts fully exposed on the top layer — no shared edges, no neighbours leaning on them. Take five to eight picks per 3D capture, then re-scan when match scores drop or a layer is cleared. You cut 3D captures per part fivefold and gain the repeatability of picking from a known surface.

A flip station. A random bin hands you the blank face-down about half the time. A fixed flip table mid-cell converts both orientations, so every pick is usable instead of half of them; add one whenever the process needs a defined face or a specific hole pattern upward.

The single-part check. Between pick and place, one 2D image of the gripper counts parts. A 5 MP smart camera such as the MV-SC6050C-08M-XBN does this comfortably at line rate, feeding a reject trap before a double reaches the fixture.

Approach direction comes from the cloud, not the teach pendant: descend along the measured surface normal with a few degrees of tilt to slip under an overlapping neighbour, the technique detailed in variable pick approach offsets. Choosing between learned and rule-based checks on the verification camera is its own decision — see AI versus rule-based smart camera inspection.

Standoff, occlusion and mounting

Bin walls shadow the far corners unless the sensor is high enough. For the 600 × 400 mm bin: the far corner sits 360 mm out on the diagonal from the bin centre, and the wall top is 200 mm high at 300 mm out. A camera at height H over the centre sees that corner along a line passing the wall top at H × (360 − 300) / 360 = H/6. For that to clear the wall, H/6 ≥ 200 mm, so H ≥ 1.2 m above the bin floor. Deeper or wider bins need proportionally more height, which is why the standoff budget belongs in the cell layout, not after it.

Mount the sensor on a rigid overhead frame when the robot changes tools, or on the wrist when you accept a calibration check after every tool change — and keep the camera outside the wrist's collision envelope either way. The 3D camera working distance and positioning guide covers the placement rules; the FOV-versus-reach trade-off is the same arithmetic used when sizing boxes for robot palletizing. Fixture the bin, or plan to find it in the point cloud every cycle — a bin that shifts 30 mm between pours eats your FOV margin.

When a feeder is the better answer

Say it plainly in three cases. Parts under 1 mm thick: the layers sit inside the noise band, and thin parts also stick, so the gripper brings doubles regardless of what the cloud says. Rates above roughly 1,500 parts per hour: one robot at 5–8 s per part tops out near 450–720 per hour, while vibratory and flexible feeders run 30–120 parts per minute. Parts that nest or weld themselves together in the tote: no sensor fixes a mechanics problem.

Between those cases there is a middle path that skips 3D entirely: pour parts single-layer into a shallow tray and let a 2D camera grid-pick them, re-imaging as the tray empties. It is a simpler cell, and it works when the parts genuinely never stack.

Checks before you order

• Flat-plate noise test. Scan a matte plate at working standoff; the RMS height noise must come in at one fifth of part thickness or less. The procedure and its failure modes are in flat-plane validation of a depth camera.
• FOV and occlusion arithmetic against your bin's real wall height, as worked above.
• Cycle budget — capture plus match plus robot move inside takt, with a stated re-scan policy.
• Double-pick detection and a flip station in the layout, not added afterwards.
• Robot reach and payload with the sensor carried or frame-mounted; read the robot-side bin picking 3D vision technical requirements before you fix the arm.
• Trial pour acceptance in writing — height steps visible, match rate and double-pick rate over a timed run.

Next step: size the sensor in the camera finder, run the full cell through the system configurator, or ask us to arrange a trial pour — every order carries free expert support, and more 3D vision guides sit under the same topic.