Tubing, catheters, hose stubs and wire are the parts that break a standard bin-picking recipe. Each strand is a few millimetres across, the strands cross and knot instead of piling, and one bad grip crushes the part or drags half a dozen neighbors out with it. This guide gives you the three numbers that decide the cell: the 3D point spacing that separates neighboring strands, the straight run and clearance a grip needs, and the grips per snapshot that set the hourly rate. Along the way you get the sensor choice, the staging geometry, a worked resolution example and a parts list.
Why thin flexible parts break the usual bin-picking recipe
A rigid part hands the robot one pose, and the cell is a locate-and-move problem. A layer of 5 mm tubing hands you a hundred independent curves, each with its own axis, its own contact points and its own neighbors. Three properties drive every decision that follows.
Crossings hide the lower strand, so the camera must resolve which strand is on top, not just where strands are. Flexibility means there is no part-level pose to match — each grip gets its own axis, fitted from the points along that strand. And entanglement means extraction disturbs the pile, so anything you plan to grip must be validated before the robot moves, in one snapshot.
The same staging-first logic that works for flat parts in a bin applies here, and picking non-planar parts with 3D vision guidance covers axis fitting for rigid curved parts. Tubing adds the two requirements those parts do not have: gap resolution between strands, and a gentle, compliant grip.
The resolution rule: point spacing before megapixels
Decide the point spacing first; the pixel count and the field of view follow from it. Work from the strand diameter, then prove the result on the gap.
Target a lateral point spacing at or below a quarter of the strand diameter: 1 mm or finer for 3-6 mm tubing. Then apply the gap test — every gap that matters must hold at least two depth points. Strands resting against each other with a 1 mm gap therefore want 0.5 mm spacing; where a miss only costs a skipped grip, 1 mm spacing is enough. Set depth resolution to 0.5 mm or better as well: against a 5 mm strand that is a 10:1 margin for reading which strand sits on top of a crossing.
Worked example. Suppose the staged layer is 480 mm wide and 360 mm deep and the depth map is 1280 × 960 points:
- 480 / 1280 = 0.375 mm per point, and 360 / 960 = 0.375 mm per point.
- A 5 mm strand spans 5 / 0.375 ≈ 13 points across its width.
- A 1 mm gap holds 1 / 0.375 ≈ 2.7 points, which passes the two-point test.
- A 40 mm grip window carries 40 / 0.375 ≈ 107 points, enough to fit a grip axis with sub-millimetre residual.
Every one of those numbers scales with the field of view, not the camera price: halve the field and you halve the spacing at the same pixel count. The calculator below runs the same division for your own field width and spacing target.
Tuning the three axes independently is its own subject; 3D point cloud x y z resolution settings without distortion covers what the per-axis resolution settings actually change.
Sensor choice: snapshot camera, laser profiler or line scan
Three routes cover thin strands, and they differ in how they pay for resolution: coverage, scan time or belt motion. Specular metal is the other hard bin — that one is a lighting problem, covered in bin picking reflective metal parts; here the geometry, not the surface, sets the requirements.
| Route | How it sees | Typical lateral spacing | Choose it when |
|---|---|---|---|
| Snapshot structured light | One coded exposure covers the whole layer | 0.3-1 mm at 0.4-0.6 m working distance | Random orientations, two or more grips per cycle |
| Laser profiler, scanned | One laser line at a time | 0.02-0.1 mm across the line; pitch along the path = speed / rate | Finest edge detail, or a robot-mounted sweep of a fixed tote |
| Line-scan 2.5D on a belt | A profile per encoder step, continuously | Sub-millimetre across; set by belt speed and rate | Continuous feed straight from production |
| ToF depth camera | Whole scene in one frame | 5-20 mm at room scale | Presence and volume, never strand separation |
The snapshot range assumes a depth map in the VGA-to-megapixel class over a 300-600 mm field; scale spacing linearly with your own field at a fixed pixel count.
For staged layers with random orientations, snapshot structured light is the default. The MV-DPS200P-02 projects a coded pattern and triangulates the whole scene in one exposure, which is exactly what the two-grips-per-one-snapshot plan needs.
- Type
- Stereo / structured light
A laser profiler trades coverage for detail. The MIS-DP3850-01P reads a single laser line per shot, resolving the strand edge far finer than any snapshot depth map, and it pays for that in scan time and in precise motion while it scans. It is the route for a fixed tote swept by the robot or for the finest wall work.
- Type
- Laser profiler
On a belt that never stops, go line scan. The MV-HLVS-16K100-L60 pairs a 16K sensor line with a 600 mm emitting length and a 120 kHz sampling rate. Sixteen thousand pixels across a 600 mm width works out to about 0.037 mm per pixel, comfortably under the 1 mm ceiling.
- Type
- 2.5D inspection
Two more names for completeness. A ToF camera such as the MV-DT01WMU returns whole-scene depth in a single frame, but its lateral spacing at bin distances is an order of magnitude too coarse for millimetre strands — it measures presence and volume, not grips. When a deep tote must be swept in place, the MV-DLS300P-04 steers a stereo scan with a galvanometer; budget the same per-area scan-time penalty as any profiler.
Stage the strands: presentation and field of view
Digging into a knot is the failure mode: extraction forces tangle the neighbors, and every pull re-randomizes the pile. Stage the parts instead. A belt, shaker or inclined tray that lays the strands one to two deep turns the knot into a searchable layer, and it is what makes the resolution math above worth doing at all. Where the parts already arrive as a random single layer on a conveyor, the decision collapses to camera choice — the 2D or 3D conveyor pick-and-place comparison walks that case.
Size the field to the layer footprint plus a 10-15% margin, and to nothing more. A 400 mm belt lane with 300 mm of depth wants a 440-460 mm wide field; at 1280 pixels that is 460 / 1280 ≈ 0.36 mm spacing, so 460 mm is the number to order against. Keep the working distance between roughly 300 and 500 mm so the gripper clears the layer on approach — spacing scales with distance, so a camera sized at one distance needs rechecking if you mount it higher. Mount the camera on a frame, not the robot: a fixed snapshot position keeps calibration stable and leaves the robot's whole reach for the grips.
From point cloud to grip axis
There is no pose to match, so the pipeline finds features per strand:
- Fit a reference plane to the stage surface; re-fit it after a belt change. Fitting a plane to a point cloud shows the fit and how to read its residuals.
- Threshold the height map at 2-3 mm above that plane. Everything left is strand.
- Trace each strand's centerline through the mask, holding continuity through crossings by following the higher strand in Z.
- Slide a 40 mm window along each centerline and fit a straight line to the points inside. Accept the window when the fit residual stays at or below 0.5 mm.
- Check clearance: both neighboring strands must sit at least 2 mm away across the whole window. At 0.375 mm spacing that is a 2 / 0.375 ≈ 5-point moat around the grip.
- Confirm the candidate is topmost — no crossing strand passes above it inside the window.
- Rank accepted windows by robot travel and send the two nearest to the controller.
The 40 mm run and the 2 mm clearance are gripper numbers, not camera numbers: they come from the finger width and the compliance you build in, so set them from your gripper drawing before you tune the vision. Teaching works the same way as for any long thin part — by feature, not whole part, as in teaching long thin parts to a vision system by feature. If thin clouds make the centerlines unreliable, why thin point clouds fail in 3D matching is the diagnosis to run.
When no window passes, skip the cycle. Nudging the pile to force a grip costs more in disturbed layers than the skipped cycle costs in time.
Gripper and cycle time: two grips per snapshot
Pair the vision with two thin, parallel fingers and compliant pads. On 3-6 mm tubing, fingers close on the 40 mm straight run without collapsing the bore; vacuum cups suit larger hose but flatten thin-wall stock. Build compliance into the fingers so a half-millimetre positioning error deforms the strand instead of fighting it.
The cycle is where the money is. Budget one cycle like this:
- Snapshot 0.35 s + processing 0.25 s + approach and grip 1.4 s + insert and return 1.2 s = 3.2 s per cycle.
- 3600 / 3.2 = 1125 cycles per hour.
- 1125 cycles × 2 grips = 2250 parts per hour.
- If one cycle in ten drops to a single accepted grip, the rate is 1125 × 1.9 ≈ 2140 parts per hour.
The second grip is nearly free because the camera time is shared — the snapshot and the tracing run once and serve both grips. That is the same economy as detecting multiple objects in one snapshot, scaled from a tray of rigid parts to a pair of strands. A third gripper channel lifts the ceiling to about 3375 parts per hour at the same cycle, as long as robot travel between grip points does not eat the gain.
Controller, PoE and the robot handshake
The vision load is modest — centerline tracing is CPU work — so the buying decision is ports and expansion. The camera wants a GigE port, a verify camera and I/O want more, and if segmentation later moves to a learned model, a GPU goes in over PCIe, so keep a slot free.
The entry unit pairs a Core i3-8100 with 8 GB of RAM and a 128 GB SSD and includes a VM6200Pro license; the MV-VC3301H-128G66 covers a single-camera cell on that spec. The MV-VC5525-256G70-NN steps up to a 12th-gen Core i5, 32 GB of RAM with a 256 GB SSD, PoE camera ports and 8 built-in light channels, which carries the camera over a single cable to a gantry mount and drives a ring or bar at a verify station without a separate light controller.
- LAN ports
- 5+ LAN ports
- USB 3 ports
- 3-4 USB 3 ports
- LAN ports
- 5+ LAN ports
- USB 3 ports
- 3-4 USB 3 ports
Compare the two on the vision controllers page and pick by port count and PoE before you size anything else. Before the first cycle runs, settle the robot-side handshake — hand-eye calibration, grip-frame conventions and the fault paths for a rejected grip. The bin-picking 3D vision technical requirements checklist is a good pre-commissioning list for exactly that.
When the usual answer is wrong
Deep-bin picking of a knotted mass is the first thing to rule out. If the parts arrive as a loose tangle in a tote, budget for a staging step before you budget for a better camera; no sensor resolves a grip through a knot.
Flip side: if staging gives you a true single layer with decent contrast, 3D is overkill. A 6 MP mono smart camera with an 8 mm M12 lens over the same lane locates strands in 2D and closes the loop faster; the smart cameras category is where that comparison starts. 3D earns its cost exactly when crossings and overlaps make one image ambiguous.
A wrist-mounted snapshot only works once the strands are separated — the limits in wrist camera bin picking: when one snapshot works bite twice as hard when parts overlap, because the gripper blocks the very view that finds the grip.
One caveat to test before ordering: clear or translucent tubing passes a projected pattern through its wall and returns dropouts instead of a surface. Shoot a sample at the working distance and check that centerlines come back continuous; if they do not, a laser profiler's thin line is the usual recovery.
Parts list for the cell
- Snapshot 3D camera such as the MV-DPS200P-02, or the MIS-DP3850-01P profiler if you chose the scanning route
- Staging conveyor or shaker that lays strands one to two deep, with side guides sized to the layer footprint
- Dual-channel finger gripper with compliant pads, finger span matched to a 40 mm straight run
- Vision controller with PoE camera ports and a free PCIe slot, such as the MV-VC5525-256G70-NN
- Shroud or enclosure panel that kills ambient light over the stage
- Optional 2D verify camera after the place station, to confirm each tube left the gripper
Checks before you order
- Gap test: lay two 5 mm rods 1 mm apart at layer height. The depth map must show at least two points inside the gap; if it shows one, shrink the field of view before you blame the algorithm.
- Step test: a 0.5 mm step wedge under the camera must split cleanly into above and below along its edge.
- Yield test: run 50 snapshots over staged parts. At least 40 must offer two accepted grips, or the throughput arithmetic does not hold and either the staging or the spacing has to change.
- Cycle rehearsal: 20 consecutive cycles at an average of 3.2 s or less, with both grips placed each time.
- Clear-part test: for translucent stock, shoot samples at the working distance and confirm continuous centerlines.
Next step
Pick the sensor first and let the spacing target drive the mount: start in the 3D sensors category, filter the camera finder by type, and check the 3D topic page for the picking guides that pair with this one. For a two-camera cell with a controller, ask for a packaged quote on the quote page or contact the applications team with your strand diameter and layer size — those two numbers size everything else.
The configurator matches a camera, lens, light, light controller and cables to each other, step by step, and lists only parts that fit what you already picked.
Questions
What point spacing do I need to separate 4 mm tubing in a 3D point cloud?
Work from a quarter of the strand diameter: 1 mm spacing for 4 mm tubing. The harder test is the gap between strands, not the strand itself — two strands resting 1 mm apart need at least two depth points inside that gap, so 0.5 mm spacing gives margin. Set depth resolution to 0.5 mm or better as well, so the upper strand of a crossing reads clearly above the lower one.
Is a laser profiler faster than a snapshot 3D camera for tubing?
Rarely at the cell level. A profiler resolves finer detail across its line, but it covers area one line at a time: a 300 mm deep pass at 0.3 mm pitch is 1,000 lines, roughly 2 s at a 500 Hz rate, and the robot or belt must move precisely while it scans. A snapshot camera reads the entire layer in under half a second and validates two grips from that single shot.
How many parts can one 3D snapshot pick per cycle?
As many as the gripper has independent channels and the segmentation accepts in one image. At a 3.2 s cycle, two grips give about 2,250 parts per hour; a third channel lifts that to about 3,375 with no extra camera time, because the snapshot and the tracing are shared. The practical limit is robot travel: distant grip points add flight time faster than the extra grip saves.
How large should the camera field of view be over the staged layer?
The layer footprint plus a 10-15% margin, and nothing more. A 400 mm belt lane wants a 440-460 mm wide field. Point spacing dilutes in direct proportion to the field at a fixed pixel count, so every oversized millimetre quietly costs gap resolution — a 600 mm field over the same 1280-pixel depth map spreads 0.36 mm spacing to 0.47 mm.
Do clear or translucent tubes work with structured light 3D?
Treat transparency as a risk to test, not a given. Clear walls let the projected pattern pass through and scatter inside, and the depth map comes back with dropouts along the strand. Shoot a sample at the working distance and check that each centerline returns continuous. If points drop out, a laser profiler's thin line usually recovers the strand where a wide area pattern fails.