A camera on a robot wrist sees a fixed patch: roughly 450 mm × 450 mm for a common integrated wrist camera at normal reach. The patch is set by the sensor and the lens, and no setting on the robot controller makes it bigger. What you can do is move the patch. This guide shows you how to cover a whole picking table with three to six planned snapshot positions, search the same taught template at each one, and drive the sequence with a counter that steps forward on a miss and resets after the last pose. You get the field-of-view arithmetic that sizes each pose, the overlap rule that stops parts falling between snapshots, the program flow with its grip check, and the point at which a fixed overhead camera beats the whole approach.
The decision: move the camera, not the optics
The patch a wrist camera sees at its search height is fixed by sensor size, focal length and standoff. For the integrated class it lands near 450 mm × 450 mm, about 0.2 m² per snapshot. A shorter lens or a larger sensor widens it, and both spend pixels per millimetre to buy area. Moving the patch costs only program structure, so that is the decision: teach waypoints that tile the table, run the same locate at each, and walk the list with a counter.
Three cases cover nearly every cell. Parts arriving in a band one footprint deep around the arm: three poses on an arc, covering about 1.3 m. A table up to about 1.3 m × 0.65 m: a 3 × 2 grid, six poses. Any plan past six poses, or any table edge beyond the arm's reach: stop teaching poses and hang one fixed camera overhead.
Two neighbouring problems deserve their own reading. Where two arms share a floor, the pose plan must also respect the cell split, and vision picking workspace limits for overlapping robot cells covers that side of it. And when a single part alone fills more than one footprint, the problem is no longer coverage: template matching a part larger than the field of view treats it.
How big one snapshot really is
Footprint scales with standoff and inversely with focal length, and ground sample distance moves with it:
Standoff buys you nothing net. At twice the height the patch doubles and every pixel covers twice the ground, so coverage and detail trade one for one. Fix the standoff where the smallest feature you must find spans at least 3 pixels, then buy area with poses, not with distance.
Worked example, using a 1/1.7" sensor of about 7.6 mm × 5.7 mm giving 4000 × 3000 pixels, behind an 8 mm lens at a 500 mm standoff:
- Patch width = 7.6 × 500 / 8 = 475 mm; depth = 5.7 × 500 / 8 ≈ 356 mm. That is 0.17 m² per snapshot.
- GSD = 475 / 4000 ≈ 0.12 mm per pixel, so a 3-pixel feature is about 0.36 mm.
- A 120 mm part spans 120 / 0.12 ≈ 1,000 pixels, far more than a template matcher needs.
The LS-CMR-0166 is a 12 MP USB3 camera with a 1/1.7" sensor and a 4000 × 3000 output, so it lands exactly on these numbers. An area sensor is 4:3, so its patch is a rectangle rather than the near-square window of an integrated wrist camera; the planning arithmetic is unchanged.
- Sensor format
- 1/1.7"
- Interface
- USB3
- Resolution
- 12 MP
- Shutter
- Rolling
Run your own sensor, lens and standoff through the calculator before you teach anything:
Space the poses so nothing falls between them
Neighbouring footprints must overlap by at least half the longest part, so a part lying across a boundary is fully inside one of the two patches. For a 120 mm part that is 60 mm, about 13 percent of the 475 mm side. The steps between pose centres follow directly: 475 − 60 = 415 mm across the table, 356 − 60 = 296 mm down it.
Order the poses serpentine, nearest neighbour first, so the arm sweeps the table once instead of crossing it empty-handed. When several zones or robots share one schedule, the same nearest-neighbour logic applies at cell level: vision target routing across overlapping robot work areas.
| Pick zone | Grid | Poses | Verdict |
|---|---|---|---|
| 1.3 m × 0.36 m band beside the arm | 3 × 1 | 3 | The classic wrist-camera case |
| 0.86 m × 0.65 m table | 2 × 2 | 4 | Comfortable |
| 1.3 m × 0.65 m table | 3 × 2 | 6 | The practical limit |
| 1.5 m × 1.0 m table | 4 × 4 | 16 | Use a fixed overhead camera |
Teach the poses
- Mark the pick zone on the table with tape and note its length and width; those two numbers and the patch size fix the whole grid.
- Jog to the first pose with the camera square to the table, no tilt, a wrist orientation you can repeat everywhere, and the standoff your lens calculation gave (500 mm in the example above).
- Capture the template here. One template serves every pose only if each shares this height within ±10 mm, which holds image scale to about 2 percent.
- Jog one step, 415 mm across or 296 mm down, and teach the next waypoint; work the serpentine so the last pose sits nearest the part exit.
- Repeat to the edges of the zone, then place a real part at the centre and at each shared boundary and confirm it is fully inside at least one patch.
- Check cable dress over the full arc last; a wrist camera that catches its own cable mid-sweep stops the cell.
Hand-eye calibration is untouched by all of this, since the camera still moves rigidly with the flange, but if you are calibrating fresh, capture and solve first: robot mounted camera calibration: capture and solve. Node settings differ by controller; a knowledgebase walkthrough of configuring camera-locate snapshot positions shows the parameters on a collaborative arm.
Drive the sequence with one counter
One integer, call it pose_index, holds the plan together. The rule that matters: advance it only on a miss. Every successful pick returns to the same pose, because parts arrive in clusters.
- Move to pose[pose_index].
- Wait out the settle, 0.2 to 0.3 s, then trigger one exposure of 10 to 20 ms with the wrist light on. Trigger only while stationary: a rolling shutter smears the frame the moment you fire during motion.
- Run the locate. On a hit, approach, close the gripper and read the grip check.
- Grip confirmed: place the part, return to pose[pose_index] and match again. Do not advance.
- Miss, or the grip check reports nothing in the fingertips: pose_index = pose_index + 1. If it now exceeds the pose count, set it back to 0 and start the next sweep from the first pose.
- Log pose_index and the match score with every pick; a pose whose hit rate decays usually has a lighting or height drift, not a software fault.
Where the match runs decides the hardware. A smart camera runs the template on the device and hands the robot a part pose over the network, which keeps the wrist self-contained. The MV-SC5060XM-08M-WBN is a 6 MP smart camera with an 8 mm focal length on an M12 mount.
- Interface
- GigE
- Resolution
- 6-10 MP
- Sensor format
- 1/2.4"
- Shutter
- Rolling
An area scan camera with the matcher in the PC leaves room to grow into measurement or OCR later, and the camera finder filters by sensor format and mount. If your cycle budget forces a short settle, a global-shutter sensor such as the MV-CU020-90GC, a 2.5 MP GigE camera with a 1/1.7" sensor and a C-mount, removes rolling-shutter shear, though not motion blur.
When snapshots stop being the answer
Time the sweep before you commit to a big grid. Budget 1.0 to 1.5 s per pose, split as 0.7 to 1.0 s of travel, 0.2 to 0.3 s of settle and about 0.15 s of exposure and match, so six poses cost 6 to 9 s whenever the table is empty. Past six poses, or with a table edge beyond reach, one camera overhead reads the whole surface in a single exposure.
The trade-offs of wrist versus fixed mounting are worth a read before you commit: eye-in-hand vs eye-to-hand camera mounting for robots. Overhead, the MV-CS200-10GC is a 20 MP GigE camera with a 1" sensor and a C-mount, enough pixels for a full table in one frame.
- Sensor format
- 1"
- Lens mount
- C-mount
- Interface
- GigE
- Resolution
- 16-25 MP
Hold it on a camera stand over the table. The LS-MNT-0040 is a machine-vision stand with a 700 mm height and a 400 × 450 mm base; past that height, plan a frame mount.
- Type
- Camera stands
Going overhead also moves lighting off the wrist, and lighting a whole table evenly is its own task: large field of view lighting. If one fixed camera must serve several zones in turn, the coordination cost lands where one camera for multiple bins: resolution and coordination trade-offs puts it. And if parts arrive stacked or jumbled rather than lying flat, no number of flat snapshots helps; 3D bin picking for flat parts: choosing a vision approach draws that line.
Checks before you release the cell
- Grid target: lay a printed grid across the zone, visit every pose, and confirm the shared edges overlap by the planned 60 mm.
- Resolution: place the smallest feature at the far corner of the farthest patch; it must still span 3 pixels, about 0.36 mm in the worked example.
- Height: measure camera-to-table distance at the first and last pose; both inside ±10 mm of the taught standoff.
- Counter dry run: cover the lens and let the program sweep; it must visit every pose, reset to the first and never fault.
- Grip-check test: present a part 30 mm off its taught pick point so the grasp fails; the program must treat the empty gripper as a miss and advance.
- Cable and light: run the full cycle at production speed and watch the wrist cable and the light uniformity at the frame edges.
When the plan is set, the camera step of the configurator sizes sensor and lens against your table numbers. If the pose grid fights your arm's reach, send the zone dimensions through the contact page and we will check the arithmetic with you.
Questions
How many snapshot positions does a 1.2 m by 0.8 m table need?
With the 475 mm × 356 mm footprint from the worked example and a 60 mm overlap, the steps are 415 mm and 296 mm. Along 1,200 mm you need ceil((1200 − 475) / 415) + 1 = 3 poses; along 800 mm, ceil((800 − 356) / 296) + 1 = 2. That is a 3 × 2 grid, six poses, and it sits right at the practical limit where a fixed overhead camera starts to win.
Would a shorter lens give me the same coverage with one snapshot?
Half the focal length, 4 mm instead of 8 mm, doubles the patch to about 950 mm × 710 mm at the same 500 mm standoff, and halves the detail to about 0.24 mm per pixel. Wide lenses also pull distortion into the frame edges, where template scores drop first. Go wider only while your smallest feature still spans at least 3 pixels; past that point, poses are cheaper than optics.
Does every snapshot position need its own template?
No, if you hold every pose at the same standoff, within about ±10 mm, the part's image scale shifts by roughly 2 percent, which a template matcher absorbs without retraining. Give a pose its own template only when it must sit at a different height, for example to clear a fixture, and capture that template at that pose. One shared template, with a short list of exceptions, is the normal working arrangement.
How much cycle time do the extra snapshots cost?
Budget 1.0 to 1.5 seconds per pose: 0.7 to 1.0 seconds of travel, 0.2 to 0.3 seconds of settle, and about 0.15 seconds for exposure and match. A six-pose sweep of an empty table costs 6 to 9 seconds. Start each cycle at the pose that produced the last pick, because parts arrive in clusters, so that pose is the most likely hit and most sweeps collapse to a single snapshot.
When should I stop adding poses and mount a fixed camera?
Past six poses, or whenever the table extends beyond the arm's reach, switch to one camera overhead. A 20 MP camera on a 1" sensor, such as the MV-CS200-10GC, reads a full table in a single exposure with no sweep time at all, and you move the lighting problem off the wrist. The trade is a mounted camera, a bigger lens and lighting a much larger field.
Why keep the grip check if vision already found the part?
Vision reports a part, not a grasp. A part shifted 30 mm from its taught pick point, or a double pick, still returns a confident match, and without a fingertip or gripper input the program marks it picked and moves on, so the table never gets re-searched and the cell slowly loses parts. The grip check tells the counter to stay put after a real pick and to treat an empty gripper as a miss.