GigE camera on a robot wrist: network setup and image retrieval

A GigE camera bolted to a robot wrist will never appear in the teach pendant, and accepting that first saves the most time: the pendant and the offline programming tool configure the robot, not Ethernet peripherals. A GigE Vision camera is a network device that answers to whichever computer owns its subnet. This walkthrough fixes the architecture, sets the IP once, routes a cable the arm can flex for months, and pulls frames into PC software that hands coordinates back to the robot. Plan one bench afternoon for first light and a day for the arm-mounted build. If wrist versus overhead mounting is still open, settle it with the eye-in-hand versus eye-to-hand comparison before buying anything.

Why the pendant cannot see the camera

Robot controllers ship with vision interfaces for their own camera families, wired into their own image service. A standard GigE Vision camera speaks an open protocol stack instead: control commands ride the GigE Vision Control Protocol (GVCP) on UDP port 3956, and pixels stream with the GigE Vision Stream Protocol (GVSP). Unless the controller maker documents support for that stack — a few do, most do not — no pendant menu will ever produce a frame, and hours spent searching its dialogs are wasted.

Give the camera a computer that speaks its language: a vision PC, or an industrial controller running a GenICam stack. Connecting a GenICam camera to an industrial controller over Ethernet covers what the host must provide, and what to check when the host is a PLC controller does the same for a PLC host.

The architecture that works

The layout that survives production has four pieces:

• One camera on the wrist, powered, focused and locked at its working distance.
• One high-flex Ethernet cable down the arm to a junction at the base, then a fixed run to the vision PC.
• A vision PC with two network ports: one dedicated to the camera link, one on the plant network that also reaches the robot controller.
• One TCP message path carrying X, Y and angle from the PC to the robot. Images never enter the robot at all.

The dedicated port is not fussiness. At snapshot sizes a single camera can fill a gigabit link, and mixing it with plant traffic is where lost frames and latency spikes begin. GigE camera network planning: switch ports and bandwidth extends this to multi-camera cells.

Latency to plan around: a 2.5 MP camera delivers a software-triggered frame in 30-60 ms including a 5 ms exposure, while a 12 MP frame needs about 110 ms on the wire alone. Budget 150-250 ms from trigger to coordinates in the 12 MP case and 50-100 ms on the small sensor, plus your processing time. For the capture code itself, follow the GigE Vision capture and save walkthrough.

Pick a camera the wrist can carry

Mass, connector security and shutter type are what matter on a wrist. A C-mount camera plus a compact lens sits far below any wrist payload that can also hold a gripper, and screw-locked connectors survive vibration that pops friction-fit plugs loose. Shutter is the real decision: stop the wrist for every snapshot — the normal pattern — and a rolling shutter is fine and gives the most pixels; strobe the light or shoot on the move, and you want a global shutter.

The MV-CU020-90GC is the global-shutter end of that choice, at 2.5 MP with a 1/1.7-inch sensor, C-mount and GigE. The LS-CMR-0053 is the pixel-count end, at 12 MP on a 1/1.7-inch sensor with a 9 fps rolling shutter.

Browse GigE area scan cameras or the camera finder for C-mount GigE models between those two sizes.

Set the IP on the bench, once

Everything here happens on a bench, before the camera touches the arm:

1. Power the camera. Most industrial GigE cameras take 12-24 V DC on the I/O connector; PoE versions are the exception, so read the power input line before wiring.
2. Patch the camera into a dedicated NIC on the PC, on a subnet nothing else uses.
3. Give that NIC a static address: 192.168.100.5, mask 255.255.255.0, no gateway. The camera link is a dead-end network on purpose.
4. Open the GigE Vision configuration tool that ships with your SDK. A camera with no address assigned self-assigns a link-local 169.254.x.x value after a few seconds — opening a link-local camera by IP from Linux shows that state. Write a persistent IP of 192.168.100.10 to the camera and power-cycle it. Persistent means it returns on that address after every reboot, which is what an unattended cell needs.
5. Ping the camera, then set jumbo packets to 9,000 bytes on the NIC and on the camera through the same tool. On 1,500-byte packets you lose a quarter to a third of usable throughput to per-packet overhead.
6. Open a GenICam viewer, set exposure to 5 ms and take one frame. The histogram should sit mid-range under your production lighting; setting exposure time through a GenICam API is the programmatic version of this step.
7. Focus at the working distance you plan on the arm, then tighten the focus lock ring.

Do all of this on a stand rather than in your hand — the camera stands category carries bench options. The LS-MNT-0040 is a slide-table stand with a 700 mm height and a 400 x 450 mm base, and the LS-CBL-0117 is a 15 m locking RJ45 cable for the bench run to the PC.

Cable the moving arm

The cable is the entire reliability problem of a wrist camera. An office patch cord fails in weeks of continuous flex; specify a drag-chain rated cable with screw-locked connectors, strain-relieved at both ends, and add a service loop at the wrist so the bend point wanders instead of fatiguing one spot.

The LS-CBL-0199 is the pattern, a 3 m high-flex drag-chain GigE cable with screw locks. Three metres covers a compact collaborative arm routed through its dress channel; on a larger six-axis arm take a longer length from the same high-flex range in camera cables, and count the whole run — arm plus fixed section — against the 100 m GigE segment limit.

If the camera is not PoE, its power follows the same path: route a 12-24 V pair through the dress pack to the I/O connector. One cable down a wrist is routine; two is where connectors start working loose, so put them on the maintenance schedule.

Bring it up on the wrist

8. Mount the camera on a rigid bracket at the working distance you focused for, and re-check focus on the arm — a lens set for 250 mm is visibly soft at 150 mm.
9. Route the cable with strain relief and the service loop, then run the shake test above.
10. Re-open the viewer with the arm moving: full frames, zero resends, steady frame rate.
11. Solve hand-eye calibration next. Robot mounted camera calibration: capture and solve is the procedure, and eye-in-hand calibration pose selection lists the robot positions that condition the answer. Record the residuals; they are your reference when picks drift.
12. Wire the coordinate path: the PC opens a socket to the controller and sends X, Y and angle per pick, as in sending camera part coordinates to a robot controller over TCP.

When picks later land off target, work the eye-in-hand pick point offset diagnosis before touching focus or IP settings. It is usually calibration pose coverage or a bumped camera, not the network.

Bandwidth: the one-camera-per-port rule

Gigabit Ethernet carries about 940 Mbit/s of payload with jumbo packets. Hold a link under roughly 80% of that — 750-800 Mbit/s — or the resend machinery starts masking losses and latency climbs. Mono snapshot cameras share a switch port happily; a 12 MP camera at full rate gets its own port.

Color triples the bytes per pixel, so 2.5 MP color at 10 fps is 600 Mbit/s and already deserves a dedicated link. A wrist rarely needs full rate: one snapshot per pick at 3 picks per second is 3 fps, where even the 12 MP camera draws 288 Mbit/s.

Field of view, pixel size and exposure at the wrist

Worked numbers for a grasp region at a 250 mm working distance. Take a 1/1.7-inch sensor: about 7.4 mm wide, roughly 4,000 pixels across, so 7.4 / 4000 ≈ 1.85 µm pixels. With a 12 mm lens:

FOV ≈ 250 × 7.4 / 12 ≈ 154 mm — call it 150 mm, a two-part grasp region. Across 4,000 pixels that is 154 / 4000 ≈ 38 µm per pixel in object space, so a 1 mm feature spans 26 px and sub-millimetre offsets are measurable. The sensor width and pixel size are typical values for the format; both formulas are the thin-lens and geometric-depth estimates.

Depth of field, at magnification m = 7.4 / 154 ≈ 0.048 and a circle of confusion of two pixels ≈ 3.7 µm: f/5.6 gives about 19 mm, f/8 about 27 mm. Part-height variation of ±10 mm sits inside f/8, so set the aperture there and leave it.

Motion is the exposure decision. Stop-and-shoot at 5-10 ms with the wrist stationary is the standard pattern and needs no strobe. Shooting a part still moving at 60 mm/s, one pixel of blur — 38 µm — builds in 0.6 ms, so a 5 ms exposure smears 300 µm, about 8 px. Either stop the wrist or strobe: a 20 µs flash at 60 mm/s smears 1.2 µm, which is invisible.

When the wrist is the wrong place

If the dress routing fights the cable at every re-dress, or duty runs around the clock, put the camera overhead and keep the arm clean. Covering a whole table with one wrist camera means touring snapshot poses, as in multiple wrist camera snapshot positions to cover a large table, or going to fixed cameras. When cycle time rules, remember each wrist snapshot includes arm travel to the pose; an overhead camera shoots during the previous pick. For a full inspection cycle built this way, see planning a collaborative robot vision inspection cycle.

Checklist before the first pick

• Persistent IP written to the camera and surviving a power cycle
• Host NIC static on the camera subnet, jumbo packets at 9,000 bytes on both ends
• Host firewall open for UDP 3956 and the GVSP stream ports on the vision NIC
• Drag-chain cable with screw locks and a service loop; shake test passed with zero resends
• Focus locked at the true working distance; aperture at f/5.6-f/8 for part-height variation
• Exposure fixed: 5-10 ms stopped, or a strobe under 50 µs if anything moves
• Hand-eye calibration solved, residuals recorded
• Socket message format agreed with the robot program, handshake included
• Trigger-to-coordinates latency measured and inside the cycle budget

With the camera chosen, configure a system walks the full parts list, and the contact page reaches us for a cell layout review.