Humanoid robot inspection sensors: the two-tier answer

A humanoid that only ferries totes is a truck with legs. The deployments that stand out give the robot a second job: it roams the floor, finds a part or a panel, and judges what it sees. That is two very different sensing problems stacked on one carrier. Finding a target from meters away tolerates tens of millimeters of position error; judging a defect does not. This page sizes each tier with numbers you can check: how depth error grows with range, how many pixels a feature needs, and how to turn that into a stand-off distance. A worked example sizes a two-tier kit for a bracket 5 m out, and the closing checks are what to confirm before ordering.

The decision: two tiers, one job each

Give the humanoid two sensor tiers with separate jobs and one handshake. The long-range tier is a 3D sensor: it detects the target, estimates its pose, and keeps the platform out of trouble while it walks. The short-range tier is a high-resolution 2D camera the arm carries to the part. The handshake is a single inequality: the position error the 3D sensor leaves behind must fit inside the capture field the 2D camera sees when it stops.

Spec both tiers from the top down. Take the smallest feature you must judge, divide its width by four, and you have the largest pixel the judging camera may have at the capture stand-off. Then require the 3D sensor's error at patrol distance to be a small fraction of the field that stand-off produces. Both steps are arithmetic, and the worked example below runs them end to end.

Tier 1: find and approach with a long-range 3D sensor

On every patrol stop the long-range tier answers three questions: is the target where the route expects it, what pose is it in, and is the path still clear. The number to watch is percent-of-range depth error, because it scales with distance. A sensor that holds ±1% of range misses the bracket's true position by 0.01 × 5000 mm = 50 mm at a 5 m stop, and by 100 mm at 10 m. That is approach-planning grade: good enough to walk in and aim, useless as a measurement.

Depth accuracy at range is mostly geometry: the wider the stereo baseline at a given distance, the smaller the depth uncertainty, which the long range stereo depth accuracy with a wide baseline guide works out in full. When you read sensor sheets, keep precision (repeatability, shot to shot) apart from trueness (correctness against a reference); the 3D camera specs guide untangles that vocabulary.

Person detection and free-space awareness sit in the same tier but are specced separately from inspection: the safety layer wants wide coverage and constant refresh, the inspection sensor wants accuracy in a narrow cone. If the navigation set is yours to choose, the guide on selecting SLAM sensors for a budget mobile robot walks the lidar-plus-SLAM trade-offs.

The find-and-approach job maps to these families in the shop:

A projected structured-light camera like the MV-DPS200P-02 suits mid-range pose on parts and fixtures. A laser galvanometer stereo camera like the MV-DLS300P-04 steers a scan across larger surfaces, which is the door-and-panel case. A ToF camera like the MV-DT01WMU trades detail for wide-area depth, which suits navigation and person awareness. Laser profilers resolve the finest cross-sections but need a sweep; on a humanoid, the arm provides it. The full range sits in the 3D sensors category.

Tier 2: judge with a close-range camera and the 4-pixel floor

The judging tier is a 2D camera that stops close to the part. Its sizing rule comes from character recognition, where the smallest feature of a character is imaged with at least 4 × 4 pixels before OCR reads reliably: put at least 4 pixels across the smallest feature you must judge. Below the floor the feature dissolves into noise. Well above it, sub-pixel edge fitting locates an edge to a fraction of a pixel under clean optics, which is what turns detection into a measurement.

Run the rule backwards to pick the camera. Feature width times the pixels you want across it sets the required pixel count; dividing the field by the pixel count gives the largest field of view that keeps the floor. The calculator does that in one step, and its button filters the shop to cameras with enough pixels:

Shutter choice follows the capture procedure. If the robot stops and settles before every capture — the normal inspection routine — a rolling shutter is fine. If it must expose while walking, a rolling shutter shears the frame and motion blur eats pixels; pick a global shutter and hold the exposure under half a pixel of travel. At 0.5 m/s with 0.04 mm pixels, half a pixel is 0.02 mm, so the exposure ceiling is 0.02 mm ÷ 500 mm/s ≈ 40 µs.

Color follows the defect: mono for shape and geometric defects, color for stains, corrosion and discoloration. The area scan cameras category filters by shutter, resolution and spectrum, and the camera finder narrows by interface once the stand-off math is done. For captures on the move, the same class exists with a global shutter in a high-speed variant:

Worked example: a bracket 5 m out, judged with two tiers

The case: a humanoid stops 5 m from a machined bracket, must find it, then grade a 1 mm wide scratch on its face — a typical cosmetic acceptance limit. The inputs, with where each comes from:

• Patrol stop distance: 5 m (typical for a roaming inspection route).
• Depth sensor class: ±1% of range (a typical mid-range 3D camera class).
• Judging camera: LS-CMR-0166, 4000 × 3000 px on a 1/1.7-inch sensor, USB3 (catalog).
• Sensor width: ≈ 7.5 mm (by the optical-format convention a 1/1.7-inch format is ≈ 9.4 mm across the diagonal; 4:3 gives ≈ 7.5 mm width — an estimate).
• Lens: 25 mm focal length (a common C-mount choice at these stand-offs — illustrative).
• Detection floor: 4 px across the scratch (the rule above).

Two rules connect the inputs. Position uncertainty equals range times relative depth error. The scratch is detectable when at least 4 pixels span it, so the widest usable field of view is the scratch width times the horizontal pixel count, divided by four.

Localization: 0.01 × 5000 mm = 50 mm. The bracket's position is known to about ±50 mm — approach-planning grade, not measurement.

Detection stand-off: FOV = 1 mm × (4000 ÷ 4) = 1000 mm. Magnification m = 7.5 mm ÷ 1000 mm = 0.0075. Working distance with the 25 mm lens: WD = 25 mm × (1 + 0.0075) ÷ 0.0075 ≈ 3358 mm, so about 3.4 m.

Grading stand-off: at WD = 500 mm the field is 500 mm × 7.5 ÷ 25 = 150 mm; the pixel is 150 mm ÷ 4000 = 0.0375 mm; the scratch spans 1 mm ÷ 0.0375 mm ≈ 27 pixels.

The decision: localize at 5 m to ±50 mm, close to under 3.4 m where the scratch first clears the 4-pixel floor, and stop near 0.5 m to grade it, where 27 pixels span the feature and edge fitting resolves its width to a fraction of the 0.0375 mm pixel. Judging from the patrol stop stays off the table: there the field is 5000 mm × 7.5 ÷ 25 = 1500 mm, the pixel is 0.375 mm, and the scratch spans under 3 pixels.

The example's judging camera is this 12 MP USB3 part with 4000 × 3000 output:

Why one sensor cannot do both jobs

The shortcut — one camera that finds and judges — fails on the same arithmetic. Judging needs the pixel at a quarter of the feature or finer. At the 5 m stop the example's camera has a 0.375 mm pixel against the 0.25 mm ceiling, and clearing the floor on a 1.5 m field needs 6000 columns — past common area sensors (a 12 MP part has about 4000 columns, a 25 MP square sensor about 5120). A longer lens shrinks the field instead: a 75 mm lens at 5 m gives 5000 mm × 7.5 ÷ 75 = 500 mm and 8 pixels across the scratch — detection, not grading, and it starves the approach tier of context.

The 3D tier cannot fill in either. Its ±1%-of-range error is 50 mm at the same stop, fifty times coarser than the 1 mm feature, and depth cameras have a minimum range — pushed close they return noise instead of depth, as the depth camera minimum range guide shows in a point cloud. Two sensors, two ranges, one clean handoff is the architecture that survives a plant floor.

The near-exception is arm-swept laser profiling: a profiler scanned across the part by the arm measures geometry at close range, but it does not catch the appearance defects a color camera judges, and the sweep still happens close, never at patrol distance.

Mounting, exposure and sync on a walking carrier

An arm-mounted camera adds one calibration task: nothing about the mount tells you the camera's pose relative to the wrist, so run a hand-eye calibration against a target after fitting it. The short-range camera hand-eye calibration guide picks the target type and stand-off for exactly this case. Re-run it after any mount, lens or refocus change; a few tenths of a degree of tilt moves the field by millimeters at a half-meter stand-off.

Write the settle into the mission: stop, let the arm's swing decay, then trigger. Exposure while settled can run to milliseconds; exposure while moving must obey the half-pixel blur ceiling from Tier 2.

Several sensors on one carrier need a common time base. The multi-camera point cloud synchronization guide handles the timestamp side, and when a wired trigger drops a frame or fires twice, the abort a capture waiting for a trigger guide shows how to stop cleanly without missing edges.

Give the captures somewhere to run on board. A vision controller — a compact vision box with camera ports — takes both tiers on the chassis, or an industrial mini PC carries the general compute. Plan the cables with it: USB3 lines stay short, while GigE reaches 100 m per run, which matters on a full-height torso.

Checks before you order

Run this list before the purchase order:

• Pixels across the smallest feature. Feature width ÷ pixel size at the planned stand-off; under 4 is a no, and grading wants several times the floor.
• Handoff margin. Half the capture field at the stop minus the 3D sensor's range error: in the worked case 75 mm − 50 mm = 25 mm per side, which is why the example detects at 3.4 m before grading at 0.5 m.
• Shutter against procedure. Global shutter and a 40 µs-class exposure if anything captures while moving; otherwise a written settle-and-stop step.
• Laser class around people. Every scanning emitter checked for swept-beam use, not just its fixed-head rating.
• Calibration budget. A hand-eye target, a repeatable fixture, and a re-run after every mount or lens change.
• Cable plan. GigE for runs past a few meters or down a torso; USB3 for short, fixed runs.

The kit in one line

Two tiers, one handoff: a percent-class 3D camera finds the part at patrol range and tolerates tens of millimeters of error, and a 12-20 MP camera closes to a stand-off where the smallest feature spans at least four pixels — for millimeter features with a 25 mm lens, roughly half a meter. Size your own stand-off in the configurator, and if the platform design wants a second opinion, the contact page reaches us. More roaming-inspection patterns sit on the applications topic page.