Vision systems for dimensional or feature checks?

Quote from Bobsmith

With that said, we are primarily looking at use Fanuc robots, likely with the R30iB control. I am curious if it is best to stick with Fanuc vision offerings and support, or if there are much better and more economical solutions? I honestly have not even learned how these vision systems are programmed, but I will learn it.

Fanuc (and this is the same for other robot brands) has a limited number of vision systems that are pre-made to work almost seamlessly with their robot operating system. The package includes both hardware and software. The vision system still has to be programmed, but it is programmed and operated through the robot's own user interface.

OTOH, you cannot simply connect any old camera (like a webcam) to a robot. The robot has no generic image interface. When you buy the Fanuc vision system, you get a dedicated hardware interface (Framegrabber, image processing, etc).

3rd-party vision systems exist (Cognex and Keyence to name two major names), but they are standalone systems, with their own cameras and other hardware. They generally have their own separate programming, software, and programming interface. And instead of sending image data to the robot, they normally communicate simple commands: "Measure this feature, feature found at X,Y with Size S", and so on.

Both approaches work. 3rd-party vision gives you more fine control, and more options for special requirements (for example, Fanuc does not sell a tiny "chapstick" camera for getting into very tight spaces). Buying Fanuc vision puts more limits on what you can do, but provides much easier integration -- hand-eye coordination being one major example. The Fanuc vision builds in tools to make that easier, while a 3rd-party vision would require you to do it all yourself.

And any vision system, regardless of brand, must be chosen based on what it's required to do. a 2D greyscale system cannot replace a 3D structured-light color system, for example. Sensor resolution, lens size, lighting, these are all critical factors. Trying to detect small defects on a bright shiny metal part requires a very different setup from simply picking parts off a clean, well-lit conveyor belt.

Quote from Bobsmith

From what I can questionably gather, these systems are examining the contrast at the pixel level to define a shape or outline, and generally if things like distance and pixel density are known, the pixels can be counted and converted to a dimension?

At the end of the day, it's all about contrast. Two adjoining pixels whose brightness values are different by more than some programmable threshold value define a step. A bunch of steps more-or-less along a straight line (again, programmable threshold) define a line. Ditto for a circle. Lines or distances between them can be measured. Circles can be measured. "Blobs" can be taught and then identified.

Things that look obvious to the human eye often are not to a vision system.

There are multiple algorithms for tweaking pixels based on their neighbors (binarize, blur, erode, sharpen, etc). And every vision system includes "finder" tools to identify features like lines, circles, text, etc.

Every vision system natively measures in pixels. Converting that to mm requires "intrinsic" calibratino -- usually a precise checkerboard placed in front of the sensor, with the sensor informed that the size of each grid square is Xmm. This also corrects for lens aberration, and often is used as part of "extrinsic", or hand-eye, calibration.

Quote from Bobsmith

I am still trying to grasp the interaction between the camera system and the robot programming. In my head, it would almost seems like tons of macro variables. "If the part is like this, do this" sort of thing.

Depends a lot on what you're doing. Typically, the robot sends a command "run Program X" to the vision system (most systems can support multiple different programs for different parts), then a "take image now" command based on the part being present, the robot out of the FOV, etc.

The vision system executes the image processing and various "finders" internally, then processes out results based on its internal program: how many parts it found, their locations in the FOV, how well they matched the "ideal" model... almost anything, depending on the system and how much processing speed/time it has. These values are sent back to the robot program, usually as internal system variables or via I/O.

Quote from Bobsmith

It certainly appears that vision in a Fanuc robot requires a vision card? I am also investigating the different vision technologies used. I am still not certain what Fanuc offers other than saying "laser vision". Seems like either structured light or laser triangulation is common.

I'm not an expert on all of them, but Fanuc has several different vision options, 2D and 3D. You usually need to order the robot from the factory with the vision system you want pre-installed. Adding them after-market is possible, but not always simple. And if you're buying used robots, some older robots may not be upgradeable to use the modern vision systems. Different options require different hookups and support hardware in the robot cabinet.

irVision is Fanuc's base vision -- mostly 2D, although it can support doing some 3D stuff by using move-shoot-move-shoot, and combining the data from the multiple shots.

3DL is basically a 2D camera with a single-line laser projector to get some depth perception.

3DV is Fanuc's mainline 3D system, using structured light. The cameras are much bulkier, and come in multiple sizes -- bigger units are necessary for greater depth, b/c you need wider-spaced "eyes" for the binocular vision to measure distance.

Quote from Bobsmith

I do see that Roboguide is setup for vision, and even aftermarket cams? At least saw a vid where a guy setup a Sony. However, I hear what is being said about possible limitations with it.

Fanuc has an optional interface for GigE cameras, but as I understand it it only works with a specific list of tested and approved cameras. I haven't used it, so I'm not sure what other limitations may apply.

Quote from Bobsmith

We also have an odd problem I want to solve. We have parts tumblers that debur parts in some media, so the parts sort of travel around in a toroid shape, in a water solution. People stand around and pick them out, then there is the problem of counting to get them all! It's a GD fiasco! There are some parts unloaders that 'sort a' work if you have the right parts, but not always. If the parts are smaller than the media, your just F'd unless you design a different unloader.

That sounds like a nightmare for vision. First, I assume the part is rolling on all 3 axes, so its facing to the camera is completely random. You would have to teach the vision system to look for all those different orientations, which could drive processing time through the roof quickly.

Then there's the fact that the parts are moving. Machine visions doesn't work on video, and it can't chase moving objects. You'll see Youtube videos of robots chasing object on conveyor belts, but those conveyors have encoders feeding realtime motion data to the robot -- the vision takes a very fast image of the part passing under the camera (triggered by a sensor), and the robot tracks the conveyor, not the object.

Then there's whatever media/fluid the parts are immersed in. Trying to get good images through that is likely to be very hard. Maybe impossible. Definitely expensive.

For either option, you have to program the vision yourself. Cognex gives much more fine-grained control, but you have to do more work to integrate with the robot. Fanuc makes robot integration much easier, but lacks the raw power Cognex offers.

Cognex is a vendor who does nothing vision, all the vision, and nothing but vision. Fanuc is a robot vendor who offers a few vision options.

Cognex "ethernet" would probably actually be either Ethernet/IP or ProfiNet. Or, possibly TCP/IP, depending. The robot would need the matching interface option. Also, Cognex offers both "Smart" cameras (In-Sight) that can work with only power and ethernet, or software engines like VisionPro (need a good computer to run it on). I generally prefer their InSight line. No stupid license issues, for one thing.

Yes, Fanuc does love to nickle&dime you on options. But if you want to ride their bus, you have to pay their fares.

Bobsmith Did you concert to use vision sensors like Keyence IV series? It is simplified system which you teach with a nominal feature and then every time it detects the feature you have high signal on DI or network protocol.

Its been sometime since I used Fanuc and their integrated iRVision cameras, but in the end I integrated a Cognex IS8402 to inspect position relative to robot user frame. I found the camera acquisition speed, resolution and pattern tools far superior. Of course with any vision application the lighting is important before you get to the vision tools.

Cognex are designed to interface directly with a robot controller or depending on application you could interface with a PLC then to the Robot IO map.

Keyence have some great options too.

From my understanding Fanuc focus on vision for picking, including 3d bin picking applications.