Pick and place machine (the internet told me so!)

[fcb]

Back a few year ago I spent well over a 1000 hours working on a low-cost P&P machine. I have also owned a couple of Versatronics machines, but now use sub-contractors. Here are some pointers:

1. Don't start like everyone else with a bunch of steppers
Tackle the show stoppers first!

2. Go round a real SMT production line
Even go and work there for a week. If it doesn't put you off - it will give you some valuable insight.

3. It's all about the feeders
I tried loose pick up from a tray - forget that, really too complex at this stage, or you end up with vibratory mechanisms to try and space out the parts and turn them over. The feeder is everything - get that right and you are 7/10th's there.  Build a 100% mechanical feeder that can be injection moulded and is triggered by the head - hint: most of the complexity is in the cover tape spool mechanism.  You must have a really cheap feeder that can stay with the reel.

4. Consider SCARA arms
This looked like the lowest cost method to get a wide reach, the maths is harder - but very little to go wrong.

5. Forget wood & plastics for the body
They are prone to warping and very temperature dependent. I even looked at using concrete and composite stone.

6. Use someone elses nozzles
Don't waste your time designing your own, there is a fair bit to it.

7. Forget doing it for $500
You'll struggle to make something usable for less than $5K, and even if you can get away with $500 for the materials, you'll never sell enough to break even let alone make a profit.  And if you can do it - the Far East will just clone it.  Or just take one the cheap Chinese machines and make them work properly.

[C]

I think most jump on hardware first for position. All with no or limited feedback.

When you place a part, you use vision for feedback. Poor hardware(your arm) gets corrected by vision. Your arm can not place a part with out your vision.

I would suggest working on vision first. When vision can tell you what to do with your arm to pick & place a part, then you have a chance of cheap hardware.





 

[free_electron]

what i would like to have is an assisted arm. kind of like a suspended nozzle with a video camera on it.
there would be a sorting box holding parts ( loose parts, no tapes ) a keypad allows me to tell the arm what box to go to. the video camera shows me what is there , i can now fine position using a small joystick or touchpad and pick the part.
the arm now comes back to hover above the last position on the board. during this travel the camera looks at the part and rotates it with smallest edge towards camera.

i have buttons to rotate left / right 90 degrees
again : jostick touchpad allows me to fine position part and a button 'places' it

mucking around with tweezers and microscopes is VERY time consuming.
having a simply mechanism that goes to the tray and allows me to pick a part ( i will do the picking using touchpad and a click ) , then spins it always in same direction. and i will do the fine positioning of placing it.

it would be nice to be able to feed a bom so on the video overlay it shows a single line : Cxxx , 100nF
i would teel it what bin it can find th 100nf part. so i dont have to deal with that.

[nctnico]

a random person on a random YouTube video mentioned that if someone can make a pick and place machine for under $200 it'll revitalize the DIY electronic hobbyists.

That is just BS and i tell you why.
1-Any P&p machine to setup reliably for one design/product costs ten times more time than to place it manually. Ergo for hobbyists that do a few pcbs a p&p machine is no improvement.

2- the XY p&p mechanics is common knowledge, tens of designs out there nothing new, openPNP sw available. The real problem is in the parts feeders. The low cost solutions are crap, the only good and reliable solutions are automatic pneumatic or electric feeders and they cost more than $100 each. A decent design you need about 50 to 100 feeders so go figure.

For small jobs, there may be scope to handle the feeder issue by using vision to pick loose parts from trays, as vision is extremely cheap nowadays.

I agree. A systems which can use components in small bins like these (with all lids open) should be feasible: .
A simple vacuum pick-up and some way to flip components which are upside down could be very usefull and it doesn't need complex & bulky feeders. With machine vision this isn't hard to do.

[mikeselectricstuff]

2. Go round a real SMT production line
Even go and work there for a week. If it doesn't put you off - it will give you some valuable insight.

Absolutely - you really do have to use a P&P to understand all the issues. P&P is not one big problem., it's LOTS of small ones, that have to ALL be right for it to be useful.
 

3. It's all about the feeders

Just like I've been saying forever, again it's multiple small problems. Even a low-end machine has to be able to do many thousands of operations with reliably minimal care & feeding, otherwise it's just a big time-suck

4. Consider SCARA arms
This looked like the lowest cost method to get a wide reach, the maths is harder - but very little to go wrong.

yes and no. Although mechanically simpler, the thing about X/Y is you can get everything off the shelf, and the inverse kinematics are trivial. SCARA has all sorts of issues like varying resolution over the working area, and the need to sometimes to go from A to B via C. Somewhere I have a video of "movement trails" made by my RV with a LED on the arm -need to find some software to process it into a timelapse that shows the flow nicely.

5. Forget wood & plastics for the body
They are prone to warping and very temperature dependent. I even looked at using concrete and composite stone.

Fibreglass (e.g. 3.2 or 4.8mm PCB material) might be worth exploring. 

If you can come up with a way of really accurately sensing the nozzle position, that would allow cheap mechanics. I think some sort of laser interferometer, or a combination of interferometry and time-of-flight might be viable

[mikeselectricstuff]

I agree. A systems which can use components in small bins like these (with all lids open) should be feasible: .
A simple vacuum pick-up and some way to flip components which are upside down could be very usefull and it doesn't need complex & bulky feeders. With machine vision this isn't hard to do.

That would be too hard - you need to get a few parts onto a flat surface and be able to shake them, so you can easily see individual parts and measure their position and rotation

[coppice]

I agree. A systems which can use components in small bins like these (with all lids open) should be feasible: .
A simple vacuum pick-up and some way to flip components which are upside down could be very usefull and it doesn't need complex & bulky feeders. With machine vision this isn't hard to do.

That would be too hard - you need to get a few parts onto a flat surface and be able to shake them, so you can easily see individual parts and measure their position and rotation

It might be hard, but if you can figure out a way to solve the pick up and selection, wouldn't it provide a fantastic solution? Those trays of closely packed passive values would avoid much of the usual configuring for each project. A massive plus for prototyping and small runs.

[mikeselectricstuff]

I suppose one way would be to pick "some" ( vacuum + oversized meshed nozzle) , drop them on a glass plate shaker for vision & picking, then put the unused ones back in the right compartment (without droppping any on the way)

[BrianHG]

I would personally first investigate the vision system I would use.  Can I properly code and interface to a moving camera on a head & correct to Z-axis errors, adapt to different height of components on my source tray?  Can I do this with relative speed?  Remember, focal length come into effect here as well as scale correction.  Yes, graphic processing power exists today to do so with ease, but, programming it isn't easy...  If I cannot work and respond with vision at my work head, everything else falls apart.

[coppice]

I would personally first investigate the vision system I would use.

I agree. I think the key to anything interesting that might be done from this point on is going to be based on the innovation that cheap vision can bring to the problem. That means your priority is to work out how convoluted a mess of parts can be sorted out by the visual system, and how effectively they can be manoeuvred by the elementary manipulators you can afford to use.

[mikeselectricstuff]

I would personally first investigate the vision system I would use.  Can I properly code and interface to a moving camera on a head & correct to Z-axis errors, adapt to different height of components on my source tray?  Can I do this with relative speed?  Remember, focal length come into effect here as well as scale correction.  Yes, graphic processing power exists today to do so with ease, but, programming it isn't easy...  If I cannot work and respond with vision at my work head, everything else falls apart.

Compared to the travel times. vision time will be pretty negligible. As vision problems go, many of those in pick/place are relatively "easy". You have complete control over lighting, well defined distances, controllable background and a very finite range of things you need to identify and measure. A RasPi camera, with the lens screwed out to get close focus, and a LED illuminator should be capable of good results.

[coppice]

I suppose one way would be to pick "some" ( vacuum + oversized meshed nozzle) , drop them on a glass plate shaker for vision & picking, then put the unused ones back in the right compartment (without droppping any on the way)

Well, if they are passives, you probably want a few of the same value across a board. Even more so if you are stuffing a repetitive panel. So, this gathering a bunch of bits, and putting surplus back, might not be as much of a drag on performance as it might appear at first sight. If the shuffling of individual bits is performed in a small scoop, the whole scoop can be moved back to the original bin, and tipped in, so I don't think the reliable return of surplus bits to the bin should be too big an issue.... as long as people don't overfill the bins.

[NickPerry]

wow, this thread is blowing up, I'm holding off on a full reply till monday when I have something to show. but reading this between breaks at work I wanted to add my 2c.

about the 'use a nozzle to grab from a bin'

there are things computers are good at that humans are terribly slow at. (like calculating the position for a SCARA arm system.)

there are things that humans are great at, but computers have a terrible time with (like recognizing that 2 over lapping parts aren't just 1 funny shaped part)

now seeing individual parts in a bin is easy for us humans, for a robot it is incredibly difficult (and slow). which is why me, Mike'selectricalstuff and those auto feeder, they take a small quantity, place it on a shaker table, and separate the pieces, this makes it much more feasible, but also more expensive & complicated. it's not just a few lines of code.

this is also why I won't be doing this for this build. (at least not until I have everything else working smoothly)

[coppice]

I would personally first investigate the vision system I would use.  Can I properly code and interface to a moving camera on a head & correct to Z-axis errors, adapt to different height of components on my source tray?  Can I do this with relative speed?  Remember, focal length come into effect here as well as scale correction.  Yes, graphic processing power exists today to do so with ease, but, programming it isn't easy...  If I cannot work and respond with vision at my work head, everything else falls apart.

Compared to the travel times. vision time will be pretty negligible. As vision problems go, many of those in pick/place are relatively "easy". You have complete control over lighting, well defined distances, controllable background and a very finite range of things you need to identify and measure. A RasPi camera, with the lens screwed out to get close focus, and a LED illuminator should be capable of good results.

This is true if you limit yourself to the kind of pick and place of yesteryear. If you want to follow something like the ASM SIPLACE route, I suspect the vision issues are considerably more challenging.

[mikeselectricstuff]

I would personally first investigate the vision system I would use.  Can I properly code and interface to a moving camera on a head & correct to Z-axis errors, adapt to different height of components on my source tray?  Can I do this with relative speed?  Remember, focal length come into effect here as well as scale correction.  Yes, graphic processing power exists today to do so with ease, but, programming it isn't easy...  If I cannot work and respond with vision at my work head, everything else falls apart.

Compared to the travel times. vision time will be pretty negligible. As vision problems go, many of those in pick/place are relatively "easy". You have complete control over lighting, well defined distances, controllable background and a very finite range of things you need to identify and measure. A RasPi camera, with the lens screwed out to get close focus, and a LED illuminator should be capable of good results.

This is true if you limit yourself to the kind of pick and place of yesteryear. If you want to follow something like the ASM SIPLACE route, I suspect the vision issues are considerably more challenging.

I don't think anyone's planning on competing with the big guys, but once you remove the requirement to do it at crazy speeds with microscopic parts, I think the pricniple could be a way to at least partially solve the problem of making cheap, useable feeders

it's not just a few lines of code.

True, but with something like OpenCV, and some effort by one person, it then becomes very cheap to reproduce. The key is finding the right balance between software and hardware - once you have parts in one layer on a plate, all you are doing is finding outlines and measuring their size, position and  and orientation, which isn't that hard.

[nctnico]

wow, this thread is blowing up, I'm holding off on a full reply till monday when I have something to show. but reading this between breaks at work I wanted to add my 2c.

about the 'use a nozzle to grab from a bin'

there are things computers are good at that humans are terribly slow at. (like calculating the position for a SCARA arm system.)

there are things that humans are great at, but computers have a terrible time with (like recognizing that 2 over lapping parts aren't just 1 funny shaped part)

You don't need to make it complicated. You know what kind (shape/size) of part is in the bin and object recognisition is far beyond it's infancy so even if parts overlap you can pick the one which is on top. The OpenCV library offers are huge amount of ready-to-go algorithms. Apply these cleverly and you can probably program this needing only a few lines of code. In addition chips could be spread out on a tray and the camera can be used to read the part numbers and select the right chip. For me a machine which can use bins and loose chips would be interesting because it means I don't need to mess with reels and feeders. It will be slower but if the machine can run on it's own I wouldn't care.

[NickPerry]

I still think you're underestimating, about 5... or was it 10 years ago now? anyways, I worked with a small team coding a program to read 'fuzzy text' it was somewhat successful, and quite slow (could take full seconds at times) and that was in a 2d space. (Think of the security questions, they are chosen for humans to be easy to read, but difficult for computers) now I admit, there are a lot of tools out there, and there may be some that can make this task easier. but I'm very skeptical. I think I'm gonna K.I.S.S it until I have a good foundation to work off of. then I may dive in (and even then, using the vibration method)

but I will have a 2 camera system for picking and placing.

[RoGeorge]

Sure, pick and place at $200 is possible, but only if those $200 are used to pay somebody else to assemble the boards.
Otherwise, with $200 in parts is hard to build even a decent 3D printer.

It reminds me of this one:
http://hackaday.com/2014/06/04/thp-entry-a-300-pick-place-3d-printer/

At $200, maybe something that moves and buzz can be built, nothing more.
For sure, not a usable pick and place machine.

OpenCV is not magic. A single frame can take seconds to process, and this happens at very low resolutions. At high resolution, it can easily go to minutes. Not all parts with the same part number are visually identical. Don't count on recognizing the parts visually. Since a human can hardly do that under a microscope, a machine vision will never do it reliable. In theory, it's easy, in practice, it will just not work.

This is not a one man weekend job.
Sure, we can talk about it and come up with crazy ideas, but that's pretty much all it can be done at $200.
 

[mikeselectricstuff]

OpenCV is not magic. A single frame can take seconds to process, and this happens at very low resolutions. At high resolution, it can easily go to minutes. Not all parts with the same part number are visually identical. Don't count on recognizing the parts visually.

You would not attempt to figure out what part it is ( at least not to start with). If you can get a bunch of known parts laying flat on a plate with the right lighting, the vision task isn't even going to take seconds. Threshold it to produce outlines, then look for outlines of the approximate size you already know the parts should be and measure the rotation. 
The only time that matters is the time to find the first suitable part to be able to pick it, and you can be looking for the next one within the image during the time the head is placing the first ( even with a head-mounted camera).

If a consumer camera can recognise and track faces in real time,  and read QR codes, finding rectangles of known size in a thresholded image should be pretty straightforward and fast.
As I said, the trick is finding the right balance between hardware and software.
 

[C]

it's not just a few lines of code.

True, but with something like OpenCV, and some effort by one person, it then becomes very cheap to reproduce. The key is finding the right balance between software and hardware - once you have parts in one layer on a plate, all you are doing is finding outlines and measuring their size, position and  and orientation, which isn't that hard.

Why measure with vision, why read text?
If the computer has reference images, a part is just a "this photo matches this reference photo".  This can also be basis of navigation and position of part.
A PC board on a table is where is the table not the table photo. This lets you find outline of PC board. You can build a photo map of bare PC board for reference.
Solder paste is again where is PC board not PC board due to paste.

A new P&P job is just using what you can get from cad and adding & comparing photo references.
 

[Kjelt]

So how are component tapes made then? I guess we need the tech that those tape component fillers have, they align the components correct side up and stick it in the tape. It looks like they do it with shaking rotating and probably vision check , we need a smaller version of this 

https://youtu.be/dbhD8JTOqf0

[BrianHG]

With <500$ of cameras & electronics & lighting & mechanics, you will be stuck with doing a lot of positioning correcting on the X&Y&Z axis at every point in the game.  Unless all your mechanics are as sturdy and good as the expensive machines which are built to BIG$, your only hope to position a QFN 0.4mm pin pitch IC from above, with a questionably centered cameras and where the IC may be on the suction head may come down to judging by vision as you lower the IC down to the PCB.  Like I said, square off your vision issues first before anything else.  Investigate as if you will only have 1 camera mounted with the head to work off of.  A second camera may be a later added feature, but if you begin relying on a multi-camera solution from the start, you are only dividing focus away from the most difficult aspect of your project.

[CM800]

Might it be worth an investigation with linear encoders and possibly DIY linear motors?

I've considered for a while, the making of linear motors using a multi-layer PCB for the motor itself and a simple row of neodymium magnets along the plate.

Linear encoders with 0.244um resolution can be had over here for 'fuck all'

https://www.rls.si/rlb-linear-component-magnetic-encoder-system

[C]

CM800

  At beginning Hard drives used optical encoders and servos to position the heads. Disk space id not increase greatly until the actual heads were the position sense.
An axis will have slop past and costs more and more to try to make slop small. Cameras are directly looking at PC board to part., Axis slop will just slow the process.

[BrianHG]

If you will be using stepper motors with your own driver, I recently done a project with this guy: A4985SLPTR-T
https://www.digikey.com/product-detail/en/allegro-microsystems-llc/A4985SLPTR-T/620-1344-1-ND/2237998

Handles all your micro-stepping, current limiting & adaptive drive power needs...