Eye-in-hand vs eye-to-hand camera mounting for robots

You have a robot arm, a camera and a part that will not sit still in one place. The first decision is where the camera lives. On the wrist (eye-in-hand) it travels with the tool and sees the part up close, but its calibration is tied to the arm's motion. Fixed to the cell frame (eye-to-hand) it watches one zone and calibrates once against the robot base. This guide gives the decision rule, the numbers behind it, two worked examples with the arithmetic, and the calibration effort each mounting costs.

The decision

Choose eye-to-hand when parts arrive in one predictable zone: one fixed camera covers the whole pick area in a single image, the calibration against the robot base holds for months, and picks land within about 1 mm. Choose eye-in-hand when targets range over the arm's full reach, when bins or the arm itself block a fixed view, or when the grasp needs refinement at close range: the wrist camera resolves five to ten times more detail and closes picks to 0.1-0.5 mm, paid for with a calibration built from robot motion and a moving cable.

The number that decides it comes first: the object-space pixel size a fixed camera achieves at the mounting distance you actually have. If one fixed camera covers the whole zone and still leaves enough pixels on the smallest feature, mount it over the workspace and stop there. If it cannot, because the zone is too wide, the mount too close or the feature too fine, the camera goes on the wrist.

The numbers that drive it

Magnification ties the geometry together: sensor width divided by the field of view you must cover. Everything else follows from it. Focal length comes from working distance, pixels per millimetre at the part set the resolution, motion sets the exposure ceiling, and the part's height spread sets the depth of field you need.

Three thresholds turn those formulas into commitments. Three pixels across the smallest feature is the floor for reliable detection; one pixel is a coin toss. Half a pixel of motion blur caps exposure whenever the arm or the part is moving. Depth of field must swallow the part's height spread, and it falls with the square of magnification: generous at 0.03×, tight at 0.14×.

The calculator below runs the same formula on your own sensor and working distance, and opens the finder at the focal length it lands on.

Worked example: fixed camera over a tray

Parts stop on a 420 × 280 mm tray, the ceiling offers 900 mm of working distance, the smallest feature to locate is 0.3 mm across, and part height varies by ±25 mm.

Start from a 1-inch sensor with 5472 × 3648 pixels at 2.4 µm pitch: 13.1 × 8.8 mm of silicon. Add margin and cover 450 × 300 mm, matching the sensor's 3:2 aspect. Magnification: 13.1 / 450 = 0.029. Focal length: 900 × 0.029 / 1.029 = 25.5 mm, so a 25 mm C-mount lens. Object-space pixel: 450 / 5472 = 0.082 mm, which puts the 0.3 mm feature at 3.7 pixels, accepted.

Depth of field for ±25 mm of height: with a tolerance circle of two pixel pitches (4.8 µm), f-number = 50 × 0.029² / (2 × 0.0048) = 4.4, so set f/5.6 and keep margin. Exposure: the tray stops, so 1-5 ms works under steady light. If the conveyor runs at 50 mm/s through the shot, the blur budget is 0.5 × 0.082 = 0.041 mm, capping exposure at 0.041 / 50 = 0.8 ms; a 500 µs strobe freezes the scene outright.

That sensor-and-mount pattern is the MV-CS200-10UM: 20 MP on a 1-inch sensor, C-mount, USB3.

Worked example: wrist camera on the final approach

Targets land anywhere in a 1.2 m cell and the gripper must align to a 45 mm part within ±0.2 mm. Close range is the wrist camera's home ground.

Take a 1.1-inch sensor with 4096 × 3000 pixels at 3.45 µm pitch: 14.1 × 10.4 mm. Cover 100 × 73 mm, giving the part, its neighbours and approach context. Magnification: 14.1 / 100 = 0.141. At 130 mm of working distance, f = 130 × 0.141 / 1.141 = 16.1 mm, so a 16 mm C-mount lens. Object-space pixel: 100 / 4096 = 0.024 mm, which puts a 0.1 mm edge at 4 pixels.

Exposure while servoing at 100 mm/s: the blur budget is 0.5 × 0.024 = 0.012 mm, so exposure stays under 0.012 / 100 = 0.12 ms. Strobe the light at 100 µs, or stop the arm and expose 2 ms under steady light. Depth of field is the tight number at 0.141×: for ±5 mm of surface height, f-number = 10 × 0.141² / (2 × 0.0069) = 14.4, meaning f/16. Backing off beats stopping down: at 176 mm the same 16 mm lens runs at 0.1×, the field opens to 141 mm, and f/8 holds 11 mm of depth. The cost is 0.034 mm pixels, a 0.1 mm feature at 3 pixels.

The MV-CH120-11UC fits the role: 12 MP on a 1.1-inch sensor, C-mount, USB3, global shutter, and the global shutter earns its keep when you strobe into a moving arm.

Rates, handshake and the control loop itself are covered in visual servoing loop design for eye-in-hand picking.

What calibration costs each way

A stationary camera over a flat scene needs a planar calibration, not a motion routine. Put a grid target in the field, collect at least 9 points across it (15 to 25 spread to the corners is better) and fit the image-to-robot-plane transform. Residuals should sit under 0.5 pixel. The grid method is walked through in robot hand-eye calibration: how nine-point alignment works, and the physical proof is in eye-to-hand calibration: verify the camera-to-base result: jog the tool to a taught dot and land within 1 mm. Size the target against your field with calibration target field of view and lens coverage margins.

Eye-in-hand replaces the grid with robot motion: 12 to 20 poses, each changing the view by at least 30° of rotation, taken about at least two axes. Under 8 poses the rotation terms of the transform turn mushy; past 30 the extra poses buy almost nothing. The pose patterns that actually work are laid out in eye in hand calibration pose selection: robot positions that work. Redo the routine after any tool change, lens refocus, camera swap or collision.

Whichever way you mount, the camera hands the robot a pick point, X, Y and rotation in a frame the controller knows, over TCP/IP or UDP. The controller-side frame bookkeeping is covered in handling user frame offsets from robot vision guidance, and when picks drift, diagnose an eye in hand pick point offset isolates the layer at fault.

Eye-in-hand vs eye-to-hand at a glance

Choosing camera, lens and mount

Both roles draw from the same shelf of area scan cameras; mounting changes the arithmetic, not the camera class. Pick the sensor first: the format and pixel count that put your smallest feature at 3 pixels or better across the field you must cover. Then shutter: global the moment anything moves during exposure, rolling acceptable for a stopped scene under a strobe.

The lens follows the sensor: its image circle must cover the format and its resolution class should meet the pixel count. Two C-mount lenses from the machine vision lenses range show the two duties: the LS-FFL-0602, an 8.5 mm, 1-inch, 10 MP lens for wide overview duty, and the LS-FFL-0567, a 16 mm, 1.1-inch, 20 MP lens with distortion under 0.3% for close work.

All 2 share: M.O.D. 0.1 m; Mount C-mount. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

The fixed camera needs a rigid home. On a bench cell, camera stands such as the LS-MNT-0040 give a 700 mm column height over a 400 × 450 mm base: coarse placement by hand, fine trim by slide.

For the wrist run, USB3 keeps a 2-4 m cable simple; GigE Vision carries 100 m and power on one line when the switch sits far from the cell, and the controller-side setup for that route is walked through in connecting a GigE Vision camera to a robot controller. Assemble the whole stack, camera, lens, light and trigger, in the system configurator, or start from the camera finder filtered by sensor format and interface.

Checks before you order

Run these gates before anything goes in the cart:

• Lens coverage: the image circle covers the sensor format; a 2/3-inch lens on a 1-inch sensor vignettes the corners.
• Wrist payload: camera, lens and connector together stay under about a tenth of the arm's rated payload.
• Cable plan: flex-rated cable for the wrist, slack for the full envelope, connector clear of the tool.
• Trigger path: a hardware line from robot I/O or the light controller, with exposure synced to the strobe.
• Focus plan: depth of field covers the height spread at an f-number the light can feed; stopping down costs light fast.
• Calibration room: the wrist can physically reach 12-20 poses around the target, and the fixed camera sees the full grid.
• Shutter: global if the arm or part moves during exposure, rolling only for a stopped scene.

When the usual answer is wrong

The fixed camera loses when a bin, pallet or the arm itself crosses the view, or when parts stack: a 2D image cannot measure stack height, and why robot vision needs depth to resolve camera coordinates explains what a 3D sensor adds at that point.

The wrist camera loses when the cycle cannot pay for refinement: stopping to shoot adds roughly 0.3-0.5 s per pick. It also loses when the tool area runs too hot, too dirty or too tight for a camera and a dress pack.

The hybrid is the standard answer for sub-millimetre picks over a large cell: a fixed wide camera finds the zone and hands off a coarse pick, the wrist camera refines the last 30-50 mm of travel. It costs a second image, 200-400 ms, and it is the architecture behind 2D vision pick and place: selecting a planar guidance system.

The short checklist

• Fixed camera covers the whole zone with 3 or more pixels on the smallest feature: go eye-to-hand.
• Zone too wide, occluded, or the grasp needs ±0.5 mm or better: go eye-in-hand, or hybrid.
• The focal length from the formula is one you can actually buy.
• Depth of field covers the height spread at the f-number you chose.
• Exposure fits the blur budget at top speed, or the light strobes.
• Calibration planned: 9-25 grid points fixed, 12-20 poses with 30° rotation spread on the wrist.
• Physical proof on file: the tool jogs to a taught point within 1 mm.