Wrist camera bin picking limits: when one snapshot works

A camera mounted on the robot wrist turns bin picking into a repeat: move over the bin, stop, capture one frame, compute offsets, pick. The engineering question is which bin scenes that loop can actually resolve. A single 2D snapshot picks randomly oriented parts reliably when they lie flat, sit apart and stay inside one shallow depth band. It fails just as predictably when parts touch, stack or stand on end, because a flat image runs out of depth information. This guide sets out the scene rules, works the standoff arithmetic for a 300 × 200 mm bin with an 8 mm lens, and matches each broken rule to the fix that restores the pick.

What one snapshot on the wrist gives you

A wrist camera works eye-in-hand: it shares the robot's coordinate frame once calibrated. Each pick then runs the same loop. The robot moves to a taught snapshot pose above the bin, stops, captures one frame, the vision software locates parts and returns offsets, and the offsets transform into robot coordinates for the gripper. The eye-in-hand calibration poses you teach decide how straight that transform holds across the whole bin, so spread the taught poses over several corners of the workspace instead of repeating one angle.

Inside that frame, a flat image measures X, Y and in-plane rotation directly. It measures no depth. The third coordinate only follows from a part's apparent size or from focus, and both drift as parts shift between layers. Every limit below grows from that one gap.

The stop itself costs little time: exposure takes a few milliseconds and a locate match runs in roughly a hundred milliseconds. Count on about half a second to two seconds added per pick once approach, settling and signalling are included, and test that against your takt. Vision supplies the offset only; the approach offset programmed into the robot decides how safely the gripper enters the pile.

The three scene rules that decide the pick

A snapshot pick survives exactly three departures from a tidy layout. Hold a photo of your own bin against each rule before ordering anything.

1. Flat. Every part presents the face the matcher was taught — its stable resting face — toward the camera. A part lying on its side shows a silhouette the model has never seen.
2. Apart. Visible background separates every part from its neighbours. The gap has to admit the gripper too: parallel-gripper fingers run of order 5–10 mm thick, so parts touching edge to edge leave neither a clean outline nor an entry path.
3. Shallow. All parts sit inside one depth band. Pixel scale grows one-for-one with distance, so at a 450 mm standoff a 60 mm spread already stretches the scale by 13%, and focus set for the top layer softens the bottom of the pile.

Lighting stability is the quieter fourth rule. A matcher scores grey-level patterns, and a window or floodlight that swings across the bin during a shift moves those patterns under it. The ambient-light fixes for pick-and-place vision — shrouding, strobing, exposure discipline — apply here unchanged.

Where each layout breaks

Touching parts. Two flat, same-material parts in contact merge into one grey-level outline. The matcher finds no clean instance, or locks onto a blend of both and reports a pose that fits neither. Even with a pose, the gripper has no gap to enter.

Stacked layers. The top part hides the bottom part's edges, so the visible pose belongs to the top while the commanded pick aims underneath. Pull one part out of a stack and the pile shifts; the next frame scores a scene that no longer exists.

Standing on end. An upright part projects a different silhouette: a cylinder on end draws a circle where the taught model expects its lying profile. Its top also sits tens of millimetres above the taught depth band, so even a correct X-Y lands the gripper short or crashing.

Deep piles. Bin walls occlude progressively. At a 450 mm standoff over a 300 mm-wide bin, the walls shadow everything below the top 60 mm, while the frame still shows part-like shapes the matcher will happily score. The picks are unreachable before the maths is wrong.

Reflective metal makes all of this worse: grey levels swing with every viewing angle, and structured-light depth turns noisy on shiny faces. The 3D methods that work on reflective metal in bins and the approach for tall, non-planar parts cover both cases. Where the parts are metallic and the pile deep, this stops being a camera choice and becomes a cell design; vendor solutions for 3D bin picking of metallic parts collects the standard architectures.

Sizing the standoff for a 300 × 200 mm bin with an 8 mm lens

One worked case pins the method down. The inputs: a kit bin with 300 × 200 mm of usable part area; parts confined to the top 60 mm of the pile, two layers of a 30 mm-tall part; and a compact smart camera with a fixed 8 mm lens over a 1/2.4" 6 MP sensor. The MV-SC5060XM-08M-WBN fits: 6 MP mono, 8 mm M12-mount lens, GigE, rolling shutter.

Two rules do all the work. Field of view width equals sensor width times standoff divided by focal length; pixel scale equals field of view divided by pixels across. The pinhole estimate behind the first rule holds to a few percent here, because the half-angle is about 18°.

Arithmetic, step by step. Sensor width: the optical-format convention puts a 1/2.4" imager near a 6.7 mm diagonal, and a 4:3 layout makes the width 6.7 × 0.8 = 5.3 mm, allowing ±10% between sensors. Standoff: WD = 300 × 8 ÷ 5.3 = 453 mm, so about 450 mm. Height check: the 4:3 sensor is 4.0 mm tall, so the frame is 4.0 × 450 ÷ 8 = 225 mm and the 200 mm bin fits with margin.

Pixels across: 6 MP on a 4:3 grid is about 2,800 × 2,100 px. Scale at the top layer: 300 ÷ 2,800 = 0.107 mm per pixel, 9.3 px per mm, so a 40 mm part spans about 370 px and a 2 mm edge about 19 px. Bottom layer, 60 mm deeper at 510 mm standoff: FOV = 5.3 × 510 ÷ 8 = 338 mm and the scale drops 13%, to 8.3 px per mm.

Decision: mount the camera about 450 mm above the top layer and one snapshot covers the bin. At 9.3 px per mm the matcher keeps the 10 or so pixels it wants across the smallest feature, and even the 2 mm edge clears it at 19 px, so resolution is not the limit here. Line of sight is: the 60 mm band costs 13% of scale and still passes, but anything the walls shadow below that band does not exist for the frame. If the pile runs deeper than about two layers, the answer is depth data, not a longer lens.

The fix for each broken rule

Bin wider than one frame. Field of view scales linearly with standoff: at 900 mm the same 8 mm lens covers 596 × 450 mm, the scale falls to 4.7 px per mm, and a 2 mm edge shrinks to 9 px, where matching starts to wander. Past that point, teach a second snapshot station instead; multiple wrist camera snapshot positions to cover a large table shows the layout and the merging of picks.

Parts touching. Restore separation or add depth. A shaker under the bin, a feeder, or a scoop-and-represent step re-singulates a pile in seconds. Otherwise, height data lets segmentation cut on the height step between parts, which stays sharp where the grey-level edge of two identical faces in contact does not.

Stacked, standing or deep. This is where 2D ends. A projected structured-light 3D camera such as the MV-DPS200P-02 images the whole bin from above and returns full six-degree poses before the robot moves. A 3D laser profile sensor such as the MIS-DP3850-01P rides on the wrist and sweeps its scan line across the pile along the approach path, building the same height map during the move. The 2D versus 3D decision for vision systems and the 3D approach for flat-part bins carry the detail.

Symmetric parts. The snapshot can be perfect and the angle still meaningless: a round or mirrored part offers several orientations the matcher cannot tell apart. Symmetric part orientation matching and rotation tolerance works that case with the tolerance the gripper can actually absorb.

One camera upgrade helps every scene above. The MV-SC6050C-08M-XBN pairs a 5 MP global-shutter sensor with an integrated multi-color light on an 8 mm M12-mount lens: global shutter forgives the residual vibration at the snapshot pose, and the camera's own light holds grey levels steady while the shop lights wander.

Choose by scene at a glance

Checks before you commit

• Measure usable part area and depth band from the bin drawing, then rerun the standoff formula with your own sensor and lens.
• Budget pixels before hardware: pixels across a feature = feature size ÷ (FOV ÷ pixels across); keep 10 or more for the features the matcher keys on and treat 3 px as noise.
• Hold a gap between parts wider than finger thickness plus clearance, or the pick collides before vision matters.
• Add the snapshot stop to the takt and re-check with the cell running.
• Teach eye-in-hand poses in several bin corners and check the offset error at the far corner before the first automated pick.
• Look at the bin during the worst-lit hour of the shift; a shroud or the camera's own light beats retraining a matcher every season.

When the wrist snapshot is still the right answer

A scene that breaks a rule can still suit the wrist camera if you change the scene instead. A tray of parts presented face-up is a one-frame pick for the whole tray, as one snapshot picking a full tray per cycle shows. A shaker re-singulates a jumbled bin. And the camera that locates the pick verifies presence afterwards — wrist camera presence detection without a separate pick is the pattern — which is the lowest-cost inspection in the cell. What vision never does is feel: hidden fasteners, hidden seams and by-feel assembly stay outside the camera entirely.

When the answer is wrong: a deep, jumbled bin of mixed parts will not be rescued by any snapshot. That job belongs to a dedicated 3D cell with fixed cameras around the bin, and the technical criteria for evaluating a 3D bin-picking system are the right reading before you budget one.

If your bin passes the three rules, the whole vision system is a compact smart camera: shortlist one in the camera finder or start at the camera step of the configurator. If it fails one rule, shortlist 3D sensors instead, or send the bin drawing through the contact page and we will tell you which side of the line it sits on. The applications topic page collects more of these cell-level patterns.