Industrial 3D cameras are a core component of any bin picking system that requires depth perception and a three-dimensional point cloud. But which kind of technology and provider is the best? The short of it is there is no universal "best". For any requirement, there are good and not-so-good options. In this piece, we'll provide an overview of 3D camera providers and how they're different.
Disclaimer: We sell an AI Vision System, which includes a 3D camera. While we're big believers of our technology, we'll also provide an objective evaluation of other providers.
A brief overview of the 3D camera market
The 3D camera market is roughly split into two types of providers.
- System vendors sell an integrated system, typically a camera, controller, and software
- Hardware vendors sell the camera and only the camera.
In most bin picking applications, system vendors are the right choice. For applications that require a high degree of customization and are staffed with experienced vision engineers, hardware vendors are the better starting point. We'll provide a brief overview of both here.
System vendors
1. Eureka Robotics
Eureka Robotics' AI Vision System is an integrated solution of 3D camera, controller and software for bin picking and vision-guided robotics.
A super-model trained on millions of parts, together with software accessible to entry-level users, reduces deployment time by months compared to conventional vision projects. The system integrates with any robotic arm and operates as an open platform: models can run locally, engineers can train new parts in-house, and no subscription fees apply.
The Eureka AI Vision System is deployed in production at Toyota, Pratt & Whitney and Subaru, and has completed more than 30 million picks.
2. Mech-Mind
Mech-Mind sells the Mech-Eye camera range together with Mech-Vision, Mech-Viz and Mech-DLK, its software for vision, robot programming and deep learning.
Ten camera models cover working distances from 300 mm to 3500 mm, and the published specification table is the most complete of any vendor in this list. A high dynamic range structured light mode on the LSR, PRO and DEEP cameras is aimed at dark and reflective surfaces, with a claimed 95 percent reduction in missing and distorted points on mirror-finish metal. The software integrates with more than 1,000 robot models.
3. Photoneo
Photoneo sells MotionCam-3D cameras and PhoXi 3D scanners, with Bin Picking Studio as the picking application on top. The hardware is also sold on its own with an SDK, so Photoneo sits in both halves of this market.
Its distinguishing feature is motion tolerance. MotionCam-3D acquires data in 10 ms at up to 20 fps and is rated for object speeds to 40 m/s, which suits picking from a moving conveyor as much as from a static bin. Bin Picking Studio takes CAD of the cell and the gripper to check for collisions before a pick is committed.
4. Keyence
Keyence sells 3D vision-guided robotics as a closed system of imaging unit, controller and setup software.
The imaging unit uses four cameras and one projector, capturing 136 images per scene, with the four-camera geometry intended to remove blind spots. Path planning, automatic robot-camera calibration and a picking simulator are included, and the system re-scans after every pick. Keyence publishes no working distance, field of view, accuracy or capture time for any model on the open web, so an evaluation runs through a sales conversation rather than a datasheet.
5. Cognex
Cognex sells the 3D-A5000 area scan camera alongside its wider machine vision software ecosystem.
The camera uses what Cognex calls 3D LightBurst to return a full field of view point cloud in as little as 200 ms, built from more than 1.5 million data points, factory calibrated and rated IP65, in standard and extended working volume variants. As with Keyence, working distance and accuracy are not published openly.
6. Pickit
Pickit sells a camera, a rack-mounted processor, picking software and robot integration as one package. It does not manufacture sensors: the current HD2 range is built on Zivid optics.
Pickit is the most transparent vendor here about real-world performance. It publishes a part picking accuracy and a part location accuracy separately for every model, and states that picking accuracy runs about 1.5 times the camera's own accuracy, measured in production environments rather than in a laboratory. Every model supports both fixed and robot-mounted installation.
Hardware vendors
1. Zivid
Zivid sells structured light cameras with an SDK and no picking layer. Detection, pose estimation, grasp planning and the robot interface are the integrator's work.
Its specifications are the most rigorous published in this market, separating point precision, planarity and dimension trueness, each with its test conditions stated, and its public knowledge base is the best technical reference any vendor in this article offers. The 2+ family shares a 5 MPx sensor, an IP65 rating, 1000 g weight and a 10 GigE interface, and needs a 10 minute warm-up before the accuracy specification applies.
2. Basler
Basler sells two families through its pylon SDK. The blaze is a time-of-flight camera at 640 by 480 px covering 0.3 to 10 m, with typical depth accuracy of plus or minus 5 mm and an IP67 rating, offered at 850 nm for indoor use and 940 nm where daylight is a problem. The Stereo ace is an active stereo camera at 2472 by 2064 px with 100, 200 or 300 mm baselines.
VGA resolution puts the blaze outside small-part bin picking. Both families suit large, forgiving picks such as boxes and sacks.
3. IDS Ensenso
IDS sells projected-texture stereo cameras in several families. The N series is built for rough environments and supports arm mounting, with the lighter N4x housings aimed at collaborative robots. The X series is modular, with a 100 W projector and working distances beyond 5 m. The C and CR series are rated IP65 or IP67, and CR models compute depth on the camera itself.
IDS publishes a reproducible Z accuracy of 0.2 mm at 2 m for large-baseline narrow-field models, and around 0.4 mm at 3.5 m for the longer ones. Per-model figures come from a camera selector tool rather than published tables.
4. SICK
SICK sells the Visionary-T Mini, a time-of-flight snapshot camera at 512 by 424 px, rated IP65, IP67 and IP69, weighing 520 g, with ambient light immunity to 50 klx.
Its resolution places it outside precision bin picking. Its datasheet, however, is the most useful document in this market for anyone with dark parts, because it publishes accuracy and repeatability against target remission, distance and temperature. That is the only quantified account of the dark-surface penalty described below.
5. Intel RealSense
RealSense sells low-cost stereo depth cameras with an open SDK. The D455 covers 0.6 to 6 m at up to 1280 by 720 px, with depth accuracy stated as under 2 percent at 4 m and a published list price of $419. The D457 is the IP65 variant.
That accuracy works out to roughly plus or minus 80 mm at 4 m, two orders of magnitude away from the structured light cameras above. It belongs to a different tier and is included for scale, and because Inbolt uses one as the standard camera in its own guidance system.
3D camera specs, explained
Field of view
Field of view is the area the camera covers at a given distance, and it has to contain the bin footprint from wherever the camera is mounted. It grows with distance, which is why it trades directly against resolution and accuracy. The usable frame is also smaller than the datasheet figure, since accuracy falls off at the edges: Zivid recommends discarding the outer 5 percent and measures its own accuracy on an 81 percent centre crop. For reference, standard Euro containers run 400 by 300 mm, 600 by 400 mm and 800 by 600 mm.
Depth range
Depth range is the span of distances over which a camera returns valid data, and it has to cover a full bin and an empty one, so a 500 mm deep bin needs at least 500 mm of usable range positioned to include both the top layer and the floor. Most vendors publish more than one distance figure, and they mean different things. The Zivid 2+ MR130 is focused at 1300 mm, optimal between 1000 and 1600 mm, and rated from 800 to 2700 mm, with the accuracy specification applying to the optimal band rather than the rated one.
Resolution
Spatial resolution is the spacing between neighbouring points on the surface of the part, and it determines whether a feature appears in the point cloud at all. Zivid publishes 0.5 to 0.8 mm as sufficient for most objects and 0.25 to 0.5 mm for small parts and fine features, noting that at 1 mm spacing a 1 cm cube carries 8 to 10 points per side. Spacing scales linearly with distance, so a resolution requirement converts into a maximum mounting height. Pickit publishes minimum object sizes per camera class instead, starting at 10 by 10 by 5 mm on its M-HD class.
Accuracy
Accuracy combines precision, the variation between repeated captures, and trueness, how close a measurement sits to the real dimension, as defined in ISO 5725. Optical 3D systems are also specified against VDI/VDE 2634, which measures probing error, sphere spacing error and flatness instead, so two vendors' accuracy figures rarely describe the same test. In a picking cell the camera is one of three terms anyway, alongside the hand-eye calibration residual that ties camera coordinates to robot coordinates and the robot's absolute positioning accuracy, which on an uncalibrated arm is considerably worse than its quoted repeatability. Pickit is the only vendor publishing the resulting derating, at roughly 1.5 times camera accuracy.
Surfaces
Material optics change what reaches the sensor. Shiny metal parts reflect projector light straight back at intensities Zivid describes as thousands of times stronger than the rest of the scene, and light bouncing between the part and the bin walls adds ripples and surfaces that do not exist. Dark parts, including many moulded plastic parts, return too little light for the signal to clear the sensor noise floor, so exposures lengthen or get stacked. SICK's Visionary-T Mini datasheet quantifies that: at 2 m, 1-sigma repeatability is about 1 mm against a 90 percent remission target and 4 mm against a 10 percent target, while accuracy stays at plus or minus 3 mm for both. Transparent parts pass light through, and no vendor publishes a measurable specification for them.
Capture time
Capture time is the interval from trigger to a point cloud the application can use, and it has to fit inside the robot cycle. Zivid publishes a budget of 700 to 1500 ms of camera time within a cycle of 5 to 15 seconds. HDR multiplies it, since a difficult scene typically needs three acquisitions, while cropping the cloud to the bin cuts it, in one documented Zivid case from 2.1 s to 1.0 s. Mains lighting sets a floor by forcing exposures in multiples of 10,000 microseconds on a 50 Hz grid or 8,333 on 60 Hz. Some vendors quote acquisition time, meaning illuminate and grab, rather than the finished cloud.
Vendor specs compared
Eureka comes first because it is our own system. Every figure is the vendor's own, and the accuracy column names the metric, because the metrics differ.
| Model | Technology | Working distance | Resolution | Accuracy, vendor's own metric | Capture time | Sold as |
|---|---|---|---|---|---|---|
| Eureka ECA-M-3001 | AI stereo, no projector | 600–1200 mm | 1440 × 1080 | Not published | 0.1–1.2 s | System |
| Eureka ECA-L-3001 | AI stereo, no projector | 1200–4000 mm | 1440 × 1080 | Not published | 0.1–1.2 s | System |
| Zivid 2+ MR60 | Structured light | 300–1100 mm, optimal 350–900 | 2448 × 2048 | 80 µm point precision; trueness under 0.20% typical, under 0.35% full temp | 25–1500 ms | Camera + SDK |
| Zivid 2+ MR130 | Structured light | 800–2700 mm, optimal 1000–1600 | 2448 × 2048 | 210 µm point precision; trueness under 0.35% typical, under 0.60% full temp | 25–1500 ms | Camera + SDK |
| Zivid 3 XL250 | Structured light, laser | 1500–4000 mm | 8 MPx | 250 µm point precision; trueness under 0.2% | 250–1500 ms | Camera + SDK, fixed mount only |
| Photoneo MotionCam-3D M | Parallel structured light | 497–939 mm | 1680 × 1200 | Under 0.500 mm camera mode, under 0.250 mm scanner mode | 10 ms acquisition | Camera + SDK, or Bin Picking Studio |
| Mech-Eye PRO M-GL | Structured light | 1000–2000 mm | 1920 × 1200 | 0.2 mm Z-repeatability and VDI/VDE at 2 m | 0.3–0.6 s | Camera + software suite |
| Mech-Eye LSR L-GL | Structured light | 1200–3000 mm | 2048 × 1536 | 0.5 mm Z-repeatability, 1.0 mm VDI/VDE at 3 m | 0.5–0.9 s | Camera + software suite |
| Pickit L-HD2-MR130 | Structured light (Zivid optics) | 800–2700 mm | 2448 × 2048 | 3.0 mm picking, 1.0 mm part location | From 200 ms | System |
| Basler blaze-102 | Time of flight | 0.3–10 m | 640 × 480 | ±5 mm typical, 0.5–5.5 m | 30 fps | Camera + SDK |
| Basler Stereo ace 100 mm | Active stereo | 0.6–2 m | 2472 × 2064 | 0.28 mm depth resolution at 1 m (quantisation step, not accuracy) | 3–10 fps | Camera + SDK |
| SICK Visionary-T Mini | Time of flight | ≤16 m | 512 × 424 | ±3 mm at 2 m; repeatability 1 mm at 90% remission, 4 mm at 10% | ≤30 fps | Camera + SDK |
| RealSense D455 | Active stereo | 0.6–6 m | 1280 × 720 | Under 2% at 4 m, so ±80 mm | Up to 90 fps | Camera + SDK, $419 list |
| Keyence 3D VGR | Structured light, four cameras | Not published | Not published | Not published | Not published | System |
| Cognex 3D-A5000 | 3D LightBurst | Not published openly | 1.5M points | Not published openly | From 200 ms | System |
| IDS Ensenso C | Projected stereo | To 3.5 m | Not published openly | 0.2 mm Z at 2 m, 0.4 mm at 3.5 m | Not published openly | Camera + SDK |
Two rows are the same hardware. Pickit's L-HD2-MR130 is built on Zivid 2+ MR130 optics, and they read 210 micrometres of point precision and 3.0 mm of picking accuracy. Zivid is quoting the sensor in a laboratory, Pickit a cell with calibration, robot and gripper error included. That is the distance between a spec sheet and a working system, and it is why a camera with coarser numbers inside better software often out-picks a better sensor.
Testing beats reading an article
Specifications narrow a shortlist. They do not settle it, because what you are buying is performance on your parts, in your bin, under your lighting.
A useful trial runs real parts rather than clean samples, including the oily and the scratched ones, and covers both a full bin and a nearly empty one, since those load opposite ends of the working range. Lighting should match the plant, because published ambient limits across this market differ by more than an order of magnitude. Three things are worth recording while it runs: capture time as configured, how many HDR acquisitions produced it, and the mounting distance. None of them appears in a demo video.
Three results are also worth separating, because a vendor will usually quote whichever is highest. Detection rate is the share of visible parts the system localises. Pick success is the share of commanded picks that arrive at the destination. Bin clearance is the share removed with nobody intervening, and it is the figure that decides whether a cell runs unattended.
Related reading: grippers and end-of-arm tooling, the bin picking field guide on cell design, the buyer's guide on vendor selection, and worked examples on metal parts and plastic parts, including dark perforated plastic on a UR5e.
Send us a bin of your parts and we return a pick video and a written verdict within 72 hours.
