Bin picking reflective metal parts: getting usable 3D data

Polished castings, machined blanks and stainless stampings arrive in a bin with the one surface property that defeats depth cameras: they reflect light like mirrors instead of scattering it. The scan comes back with holes where facets face away from the projector, saturated ridges where they face toward it, and a robot that reaches for the bin wall. This guide gives you the geometry behind those two failure modes, the fill lighting and capture settings that bring the data back, and the numbers to size the set-up. You will finish able to predict which faces a sensor loses, choose between a laser profiler and an area structured-light camera, and gate every pose on the confidence image before the first pick.

Why polished metal returns no depth

Every triangulation sensor — laser profiler, projected structured light, speckle — plus passive stereo relies on the part scattering light back to the sensor. A matte surface scatters in every direction, so some return reaches the lens from any viewpoint. A polished facet obeys the law of reflection: outgoing angle equals incoming angle about the surface normal, so the return leaves in one direction only. The bright field versus dark field decision comes down to this same law.

That single-direction return creates exactly two failure modes. A facet tilted toward the projector mirrors its light into the lens: pixels saturate and depth goes invalid — the bright failure. A facet tilted away sends its return outside the sensor entirely: pixels starve — the dark failure. Passive stereo fails a third way, because polished metal offers no texture to match on, so a stereo rig returns nothing there regardless of lighting. A bin of mixed facets turns the depth map into confetti of bright spikes and dark holes, and part-matching fails on the very top faces you want to grip. The same surface-versus-geometry battle in 2D is mapped in light geometry selection by surface and feature.

The reflection rule that predicts your holes

Set the geometry in words. Put the camera axis vertical over the bin centre and the pattern projector offset by an angle called the projector offset. A facet tilted toward the projector by a tilt t reflects its light at the offset angle minus twice t, measured from vertical — rotating a mirror by t rotates the reflected ray by 2t. The facet blinds the camera when that return falls inside the acceptance cone, the narrow band of angles around vertical that reaches the lens at the bin centre.

Everything inside the blind band saturates; everything tilted steeper than it starves. The rule also gives you the fix: rotate the projection azimuth 90° and capture again. A face that dips toward the projector in the first view dips across it in the second, outside the band. Merge the two clouds and the top face fills in.

Worked example: facets that blind a projector set 30° off the camera axis

Inputs, and where each comes from:

• Sensor layout: camera axis vertical over the bin centre, pattern projector offset 30° from the camera axis — an example single-projector triangulation layout; substitute the offset your sensor actually uses.
• Acceptance cone: ±10° about vertical, the central 20° of the frame where most picks land — an example value; a tighter lens narrows it, a wider lens grows it.
• Facet tilts: 10°, 15° and 20° toward the projector — typical draft angles on a cast face, used here as the example part.

The rule is the mirror law from the last section: the reflected ray rotates twice the surface rotation. Arithmetic, one facet at a time:

• Flat face, t = 0°: return at 30° - 0° = 30° from vertical, far outside the ±10° cone. Clean measurement.
• t = 10°: 30° - 2 × 10° = 10°. The return sits at the cone edge; pixels go marginal and start to clip.
• t = 15°: 30° - 2 × 15° = 0°. The return comes straight back into the lens: full saturation, invalid depth.
• t = 20°: 30° - 2 × 20° = -10° — 10° on the far side of the axis. The facet leaves the cone and starves instead of blinding.

The blind band runs from (30° - 10°) / 2 = 10° to (30° + 10°) / 2 = 20°. Result: every facet tilted between 10° and 20° toward the projector returns glare, steeper than 20° it goes dark, and both are holes in the cloud.

Decision: capture a second scan with the bin rotated 90° — or move the sensor head to a second station 90° around the bin — and merge the clouds. A face dipping 15° toward the projector in the first view dips across it in the second, nearly flat as the projector sees it, so it measures cleanly. Where the part cannot rotate, add HDR exposures so the band-edge faces survive and route the gripper around the band core.

Diffuse fill: dome light for singulated parts, diffused bars for the bin

A dome light wraps the part with an emitting area that returns light from a wide solid angle — cloudy-day illumination. Every facet collects light from many directions, so the camera's passive channels and the 2D texture see even brightness instead of a black-and-white patchwork. Domes tame glare on curved and wrinkled parts for the same reason; dome lights for shiny, curved and wrinkled parts covers that ground.

Size the dome by a single rule: the emitting area should exceed the largest part diagonal you need to light. The LS-DOM-0085 pairs a Ø166 mm outer diameter with a Ø116 mm emitting area. That lights a singulated part to roughly Ø100 mm with margin around it. The same fill logic carries small shiny hardware; lighting shiny metal pins without glare shows it at pin scale.

A bin is wider than any dome you want to swing over a robot cell, so ring the bin instead. Long diffused bar lights put broad-angle white light across a metre of bin from each wall, and the diffused type spreads the output so no single LED makes its own glint. The LS-BAR-1276 is a 372 x 28 mm white diffused bar in the 200-400 mm size class. Ring a mid-size bin with one bar per wall. For full-length bins, the LS-BAR-0295 stretches to 1272 x 22 mm of diffused output. The light finder filters both families by size, colour and diffused type.

Sequence the fill as carefully as you aim it. Fire it between the projector's frames or in a separate exposure; a fill burning during the coded projection paints its own bright spots into the pattern and corrupts the decode. Route it through a strobed output on your light controller so the capture owns the timing.

Capture settings that fill the blinded band

HDR first. Shoot two to three exposures per projection: the short frame keeps glare faces from clipping, the long frame lifts the scatter faces out of noise, and the merge keeps, per pixel, the first frame that did not clip. Capture time grows by one frame per added exposure, so budget it against the cycle.

Second azimuth next — the worked example's rotation. Two merged clouds beat one perfect exposure, because the blind band moves with the projector azimuth while the part stays put.

Polarization when nothing can move: cross-polarize the projector and the lens. Specular light keeps its polarization while scattered light loses it, so the crossed analyzer suppresses the mirror return and passes the diffuse component. Expect to lengthen the exposure to pay for the filters; the 2D version of the trick is in cross polarization in machine vision.

Texture-less faces starve passive stereo even with perfect fill, so give the surface a pattern it can hold. Speckle and multi-DOE projectors paint dense synthetic texture across the field, so feature matching has something to lock onto where the surface itself offers none. The 3D sensors category filters for stereo and structured-light cameras, and the volume-measurement designs built for mixed loads are exactly the mixed-bin case.

Judge none of this on the RGB view. The confidence image tells you which pixels the sensor itself trusts — decode its bit flags and mask the failures before the planner ever sees the cloud, as confidence image bit flags and masking explains. For the full polished-metal capture checklist, improve shiny metal 3D point cloud capture walks the same ground from the cloud side.

Choosing the sensor class for a reflective bin

The 3D sensors category lists 170 products: 75 laser profilers, 34 stereo and structured-light cameras, 33 2.5D inspection systems, 15 logistics and volume cameras and 13 ToF cameras. Four of those classes matter in a reflective bin:

For the sweep lane, a profile sensor such as the MIS-DP3850-01P — a 3D laser profile sensor — rides the sweep and profiles the bin line by line, so each profile carries a single, known incidence angle.

Full-field capture fits the overhead cell: the MV-DPS200P-02 is a projected structured-light 3D camera that images the whole bin per pose, with HDR and the second azimuth layered on top.

When the incidence angle itself must move between sub-captures, a steered beam does it mechanically: the MV-DLS300P-04 is a laser galvanometer stereo camera, steering the beam so each sub-capture lights the facets from a different direction.

The cell, not the catalogue, picks the class. Score cycle time, part presentation and gripper clearance first; a technical criteria guide for evaluating 3D bin picking systems walks that scoring line by line. The wider trade space across stereo, ToF and structured light is mapped in industrial 3D camera selection.

Coverage numbers: standoff, lateral spacing, depth noise

Mount height first. Fit the bin's longest diagonal inside the field of view with margin, and keep the tallest stack inside the sensor's depth range:

A field 700 mm across with a 30° half-angle lens needs (700 mm / 2) / tan 30° = 350 mm / 0.577 = 606 mm — mount the head about 600 mm above the tallest stack. The 3D camera working distance and positioning guide carries the full method, near and far limits included.

Lateral sampling then takes care of itself: 700 mm across 2,000 pixels is 0.35 mm per pixel, and gripping decisions need millimetres, not fractions of them. In a reflective bin the constraint is never sampling — it is the holes. Depth noise is the number to watch: keep it under one third of the smallest clearance feature the gripper must respect. Smooth raw depth before matching (stereo depth map noise) and set the grid resolution deliberately (point cloud x y z resolution settings).

Checks before the robot picks

Run these before the first live pick:

• Mask on the confidence image, not on colour: drop pixels the sensor flags, then measure coverage over the top faces the gripper actually touches.
• Audit the holes against the rule: if missing pixels cluster where facet tilt sits in the blind band, add the second azimuth — more exposure will never fill a saturated pixel.
• Diff the fill: one capture with fill off, one with it on; a changed valid-pixel map means the fill leaks into the projection window.
• Kill ambient competition: shade skylights and high-bay lamps. Strong sunlight into a stereo rig is a filter-sizing problem, not an exposure problem — see ND filter sizing for stereo depth cameras in direct sunlight.
• Prove repeatability: scan a fixed witness part 20 times through the full capture chain and measure the pose spread. That spread, not the data sheet's best case, is what the gripper clearance must absorb.
• Check the hand-off before go-live: 3D camera robot integration compatibility lists what to confirm between sensor, PLC and robot controller.

When the standard recipe is wrong

Some bins defeat overhead fill no matter how you sequence it:

• Deep, narrow bins with steep walls: the fill and the projector both graze past the contents. Change the presentation — tilt the bin, add a divider insert, singulate with a feeder — which costs less than any light you can buy. Approach choice by part shape is mapped in 3D bin picking for flat parts, and bin picking non-planar parts with 3D vision guidance covers the guidance side.
• Matte black parts: the problem flips from too much return to too little; black O-ring lighting covers the absorption case.
• Transparent parts: specular returns plus volumetric ghosts; detecting small transparent parts needs its own lighting logic.
• Already-singulated parts: skip 3D entirely; 2D or 3D camera for conveyor pick and place prices that decision.

Where to start

Two sketches settle most cells. Draw the facet tilt histogram you expect on the part — draft angles keep it short and flat, machined faces make it tall — and mark the blind band from your sensor's offset and cone. If the band covers the faces you grip, plan the 90° second view from day one. If it does not, a single overhead capture with HDR and dome or bar fill is enough.

Match the fill light from dome lights or bar lights, the sensor from the camera finder, or start at the lighting step of the configurator. More depth-camera guides sit on the 3D topic page, and for a cell that resists all of the above, contact us with the part and the bin.