A robot head camera must do more than produce an attractive depth map. It must see the nearest manipulation area, fit the head assembly, and provide usable images when someone operates the robot remotely. This guide helps you turn those requirements into measurable geometry and acceptance checks. Start with the task distances and viewing arrangement, then choose the camera spacing; do not treat human eye spacing as a universal depth-sensing requirement.
Separate depth sensing, viewing and mechanical fit
Stereo camera spacing for robots involves three different decisions. The optical baseline controls triangulation geometry. The displayed stereo views affect how the operator perceives the scene. The complete assembly determines clearance, mass distribution and mounting stability.
These decisions interact, but they are not interchangeable:
| Requirement | What to evaluate | What baseline alone cannot establish |
|---|---|---|
| Depth measurement | Range, disparity precision and common visibility | Accuracy on reflective, textureless or occluded surfaces |
| Direct stereo teleoperation | Operator eye separation, display projection and latency | Comfort or correct apparent scale |
| Head integration | Optical-center positions, housing dimensions and movement envelope | Mount stiffness or calibration retention |
A larger baseline generally reduces depth uncertainty when other conditions remain equal. It also increases viewpoint differences, which can make nearby objects harder to match.
Use the stereo baseline selection guide for the underlying sizing decision. Here, the additional question is whether that geometry remains useful inside a moving robot head and through an operator's display.
Write separate acceptance criteria for depth measurement and teleoperation. A successful depth test does not establish that the viewing arrangement is comfortable.
Calculate common coverage at the nearest task plane
For an ideal rectified stereo pair, depth follows the familiar triangulation relationship:
Use consistent length units for depth and baseline. This expression assumes calibrated cameras and corresponding image points; it is not a specification for any catalog product.
For parallel, identical cameras with horizontal field of view θ, the ideal horizontal overlap at a frontoparallel plane is:
The field-of-view positioning guide helps connect lens coverage with mounting distance. The figure below shows single-camera coverage; a stereo pair requires the intersection of both views.
The overlap formula describes geometric visibility, not guaranteed valid depth. Rectification cropping, lens distortion, disparity limits and occlusion can reduce the usable region further. The explanation of smaller depth-image coverage covers that distinction.
Check the nearest required working plane first. A camera arrangement that covers a distant work surface can still leave a hand, tool or nearby object without usable depth.
Work through a head-camera geometry example
Consider a hypothetical pair with identical rectified views, a horizontal image width of 1280 pixels and a horizontal field of view of 90 degrees. Compare baselines of 65 mm and 120 mm. These are illustrative layouts, not specifications or recommended operating distances for any product.
The focal length expressed in pixels is:
At a distance of 300 mm, each camera covers an ideal width of 600 mm:
This comparison gives the following ideal geometry:
| Quantity | 65 mm baseline | 120 mm baseline |
|---|---|---|
| Common width at 300 mm | 535 mm | 480 mm |
| Disparity at 500 mm | 83.2 pixels | 153.6 pixels |
| Disparity at 2000 mm | 20.8 pixels | 38.4 pixels |
The closer plane requires more disparity search. Confirm that the selected processing system supports the required range rather than assuming geometric overlap guarantees a depth result.
A first-order uncertainty estimate shows the benefit of wider spacing:
Assume a disparity uncertainty of 0.25 pixel for this calculation. At 2000 mm, the narrower baseline gives approximately:
The wider baseline gives approximately 13 mm under the same assumptions. These estimates exclude calibration error, surface effects and systematic bias. The assumed disparity uncertainty must be measured for the actual scene.
Image detail is a separate constraint. At 500 mm, each hypothetical view spans 1000 mm, giving 1000 / 1280 = 0.78 mm per pixel. A 4 mm feature occupies about 5.1 pixels. At 2000 mm, that same feature occupies only about 1.3 pixels.
Use the resolution calculator to check feature sampling independently of depth uncertainty. Neither pixel count nor baseline alone establishes reliable recognition.
Match teleoperation geometry, not just eye spacing
A baseline close to the operator's measured eye separation can be a reasonable starting point for direct stereo viewing. It is not a universal comfort setting, and it is not necessary for every autonomous perception task.
For example, a measured operator eye separation of 65 mm could justify testing a matching capture baseline first. The display still needs appropriate left/right assignment, projection geometry and stereo alignment. A matching baseline cannot compensate for vertical image misalignment or inconsistent image scaling.
Record the following independently:
- Capture baseline, measured between optical centers rather than housing edges.
- Effective image field of view after rectification and cropping.
- Displayed angular field of view and any image resizing.
- Operator display settings and viewing mode.
- Capture-to-display latency and variation during head movement.
Cropping changes the effective field of view. Resizing changes the mapping between image pixels and display angles. Neither operation physically changes the capture viewpoints.
If views are synthesized from depth to create a different virtual baseline, test missing regions and occlusion boundaries. Reprojection cannot recover a surface that neither source camera observed.
Start viewing trials with a stationary scene, then introduce slow head motion under controlled conditions. Stop a trial if the operator reports discomfort. Do not interpret a short comfortable session as proof that prolonged use will be comfortable.
Evaluate catalog parts without assuming a matched kit
Compare 3D sensors through the camera finder, but request the missing geometry before selecting a head-mounted unit. The supplied catalog identifies sensing types; it does not establish baseline, head clearance, teleoperation suitability or simultaneous stereo-image access.
The MV-DPS200P-02 is cataloged as a projected structured-light camera. That identifies its sensing approach, not whether it can deliver the left and right views required by your display. Request its documented operating volume, acquisition sequence and available outputs.
- Type
- Stereo / structured light
A custom pair of area scan cameras is another architecture to investigate. The MV-CU020-90GC is cataloged as a global-shutter color GigE area scan camera. Its catalog entry does not establish a complete stereo solution or synchronized capture compatibility.
- Sensor format
- 1/1.7"
- Lens mount
- C-mount
- Interface
- GigE
- Resolution
- Up to 3 MP
A custom pair also requires suitable lenses, a stable mount, calibration, correspondence processing and verified timing. Follow the paired-image synchronization checks before attributing moving-scene errors to baseline.
For stationary geometry experiments, investigate camera stands and positioning hardware through the accessories finder. The LS-MNT-0001 is a manual tilt stage. Confirm its mechanical interfaces, adjustment range and load suitability before incorporating it into a bench fixture.
- Type
- CCD camera stand and positioning stage (LS-MNT-S01)
These parts are investigation candidates, not a qualified head-camera kit. None of the supplied entries establishes an adjustable stereo baseline or suitability for mounting on a moving robot.
Place and calibrate the complete head assembly
Baseline is the distance between optical centers, not the overall head width. Housing walls, lens barrels and mounting interfaces can make the complete assembly substantially wider than the baseline.
Draw the optical centers, protective front surfaces, mounting interfaces and required clearance in the same layout. Check the nearest task plane with hands, tools and other head-mounted equipment in their expected positions.
Place the stereo midpoint deliberately relative to the head's rotation axes. An offset camera follows a translated path when the head rotates. That translation is real motion, not necessarily a depth-processing fault.
The robot model and image-processing chain must represent the camera-to-head transform correctly. The robot-mounted calibration workflow addresses the capture-and-solve process, while the camera extrinsic accuracy checks explain why a small reprojection residual is not the whole validation.
Calibrate the final optical configuration. Changing camera spacing, relative tilt, focus or image cropping can invalidate parts of the calibration or its application. Verify which changes require recalibration according to the selected system's documentation.
Run an acceptance test at the actual task distances
Use a repeatable sequence rather than choosing a baseline from a single attractive depth image.
- Define the task volume. Record the nearest, typical and farthest required distances. Include tool approaches and temporary obstructions, not only the nominal work surface.
- Mark common visibility. Place a target grid across each task plane. Record where both cameras see each target and where the processing system actually returns valid depth.
- Measure repeatability and bias separately. At a fixed target, collect a defined sample, such as 100 frames. Compare the mean measured position with an independently known reference and calculate frame-to-frame spread separately.
- Test difficult surfaces. Include the actual finishes and object shapes. Flat textured targets alone will not reveal all correspondence failures or edge occlusions.
- Repeat during motion. Compare stationary and moving captures. Check exposure and timestamps before changing the mechanical spacing.
- Evaluate the display separately. Record operator viewing settings, latency and observations for each layout. Keep scene geometry unchanged when comparing display configurations.
Choose acceptance limits from the task. An illustrative requirement might demand valid depth over at least 95% of a defined target grid and depth error within 10 mm at a specified plane. Those numbers are not general recommendations; they demonstrate how to make a requirement testable.
Define the denominator for coverage. A percentage of the entire image is different from a percentage of required tool-contact locations. Report invalid pixels rather than silently excluding them from accuracy calculations.
When the display combines color and depth, follow the timestamp alignment checks. A geometry adjustment cannot fix a stale color frame paired with newer depth.
Know when wider spacing is the wrong answer
Choose a wider baseline only when its depth benefit survives the complete task test. It can be the wrong choice when the nearest work area dominates, thin objects create substantial occlusions, or the head assembly cannot maintain relative camera alignment.
A wider pair also gives more different views of depth discontinuities. The foreground edge visible to one camera may hide the corresponding background region from the other. More geometric overlap does not remove that local occlusion.
For direct stereo viewing, a wider capture baseline may alter apparent depth relative to the operator's display geometry. Do not assume software scaling alone will make the presentation equivalent to another physical baseline.
When sensing and viewing requirements conflict, consider separate acquisition paths: a depth system chosen for measurement and a viewing pair chosen for operator presentation. This adds calibration, timing and integration work, so evaluate the extra complexity before committing.
The camera-to-robot compatibility checklist helps identify interface and coordinate requirements before hardware selection.
Check these items before ordering
Send a supplier or applications team a compact requirement sheet containing:
- Task distances and required coverage: include the nearest plane and its required common viewing width.
- Depth acceptance limits: distinguish bias, repeatability, valid coverage and edge behavior.
- Viewing requirements: state whether raw stereo images, rendered views or a single image are needed.
- Geometry documentation: request baseline, optical-center locations, field of view and supported disparity or depth range.
- Mechanical envelope: include head clearance, allowable mass and mounting interfaces.
- Timing requirements: specify capture synchronization, output rate and acceptable capture-to-display latency.
- Calibration responsibilities: identify who supplies calibration and how the installed assembly will be verified.
Choose the layout that passes these requirements, not the one with the largest baseline or the most human-like appearance. Begin a hardware shortlist with the camera configurator, or send your geometry and acceptance sheet through applications support when the available catalog data cannot establish a fit.
Questions
Should a robot camera baseline match human eye spacing?
Not necessarily. Matching the operator's measured eye separation can provide a starting point for direct stereo viewing, but depth sensing has different requirements. Display projection, cropping, alignment and latency also affect presentation. Select and validate the capture baseline against both the robot's task volume and the intended viewing arrangement.
Does increasing stereo camera spacing always improve depth?
A larger baseline reduces depth sensitivity to a given disparity error under otherwise equal conditions. It can also increase occlusion, reduce near-field common coverage and make correspondence harder. Evaluate valid depth and measured error across the required task volume rather than selecting the widest assembly that fits.
Where should the stereo baseline be measured?
Use the calibrated distance between the cameras' optical centers. Housing edges and mounting-hole spacing are not equivalent unless the mechanical documentation explicitly relates them. Obtain optical-center drawings or calibration information, then retain the complete camera-to-head transform in the robot's coordinate model.
Can image resizing compensate for a different capture baseline?
No. Resizing changes how captured images map onto the display, but it does not change their original viewpoints. Depth-based view synthesis can create virtual viewpoints, although it introduces its own calibration, latency and missing-surface limitations. Validate the complete presentation rather than relying on a resize factor.
What information is needed before buying a head-mounted depth camera?
Request optical-center locations, baseline, usable depth range, common coverage, acquisition timing and available image outputs. For a moving head, also request mechanical and calibration documentation. A catalog description identifying a stereo or structured-light camera does not establish that it supports direct stereo teleoperation.