Current status of my DIY pick and place machine
When I started developing the RasPiComm Plus I soon realized that I need the ability to manufacture smaller batches of our boards since we will have at least 10 different extension boards. I also wanted to stick to my 100% made in the EU principle.
Usually I would use my manual pick and place system to assemble prototypes. The boards are manufactured in Germany and sent to us. For larger quantities I would send the parts and boards to another company in Germany and let them assemble them. That costs almost €4.000 per 1.000 boards just for the assembling in the case of the RasPiComm. And it will take time, up to 4 weeks depending on their schedule. I needed a better solution and faster turnaround.
no more manual pick and place
Since I have a manual stencil printer and a really cool vapor phase soldering oven a pick and place machine is the missing piece of equipment to have a reasonable low volume production facility.
So those were my options:
Buy a new pick and place machine
I compared prices of various options. An US company sells a cheap pick and place, very simplistic design (no housing, aluminum profiles), for about $25.000. Shipping to Austria would have added another $2.500. The machine itself was not impressive at all. I had my doubts if it could handle 0402 relieably.
The TM220A, a $4.000 toy-grade pick and place as Dave L. Jones calls it (and I am completely with him on that) has no vision. And it does not even look like it is worth $4.000. Not an option at all.
Buy a used pick and place machine
Tempting. Ryan O’Hara bought a Quad PnP for about $16.500 (without the feeders). He talked about it on The AmpHour. It is not that easy if you are not living in the States. I cannot come over and have a look if this machine is what I want and shipping is of course awfully expensive to Austria. Another thing is the software. This price only applies to the DOS version. I like having control over the software and my processes.
Well this option has the bonus that I’d have to build something from scratch. Do you need more reasons apart from this one? I know, me neither.
Nevertheless it has to be clear that building such a machine is inevitably more expensive than buying one in the short-run. If you don’t consider your time worthless of course. But one gets better with each project, you will learn a lot and you will have a system which you understand down to the tiniest bit of wire and code, so you can tackle every problem that arises, you can adapt software and hardware to your needs. And as an entrepreneur I also keep in mind that it maybe will be a product someday, or parts of it, so I always consider these kind of projects as an investment for the future not only in terms of learning and training for me and my employees but also having working modules that can be reused later on.
I thought about my goals, what I needed and what I don’t need. First of all this should be a rapid prototyping project. I will make design decisions based on if they are practical and the effort/benefit ratio fits for me. It should not be a project which is never really usable. Since this is more a side project that a full-time project I gave myself 3 months. As I am writing this the 3 months are not over yet, so it is still in time.
I also did not want to think about automatic feeders too much for now. I knew I’ll have to redesign small areas of the machine for automatic feeder support. Until then working with belts held down by a polycarbonate shield should be sufficient for prototype quantities. For the small volume production target feeders are a must of course. So this is a compromise I make when starting this project – in the beginning I won’t have automatic feeders.
After some hours of 3d-modelling we milled the first aluminum parts and ordered some off-the-shelf components and aluminum profiles. Milling was done with a very simple and cheap CNC, not even my good one (I got two of them). The CNC does not have a toolchanger, not even ball screw bearings. But it does its job. After that the parts went to anodization. Black of course.
some milled pick and place parts
after black anodizing
Next step: Motors. I used quite beefy stepper motors in closed-loop mode for the X and Y axis and servos for the 2 heads I planned. Part rotation is again done by a stepper motor with a hollow shaft. Two simple turning parts made from stainless steel made the pickup base for the nozzle. I glued a magnet to the bottom. I used simple fuse clips as the nozzle holders. Nozzle-changer: done.
X and Y have ball-screw bearings with a pitch of 20mm to get the speed up. With a 500 step encoder with quadrature encoding this means a positioning precision of +/- 0.09° which would be +/- 100µm. Enough for this machine. I plan to switch to BLDCs for a couple of other reasons with a 4000 step encoder which would lower that to 50µm in normal mode and 12.5µm with quadrature encoding. This is of course hypothetical, the mechanics are not that precise but still very good since I only used high quality ball bearings for X, Y and Z axis. The guides are also precision ones.
Assembing was easy, it took a couple of hours. Wiring was a bit of a pain, it took a whole day.
pick and place partially assembled
For the vision system I am using are 2560×1680 pixel USB3.0 cameras. One from the bottom for part alignment and a top camera for PCB/feeder alignment. I made a simple ringlight with 48 650nm wavelength LEDs. Funny detail: I was not able to do a circular pattern alignment of the LEDs in Altium, so I wrote a tiny Autohotkey script which allowed me to enter a number of parts and a diameter and it automatically aligned the parts correctly. Huge timesaver when you want to play around with the radius in which the parts are aligned on the PCB.
I also added a joystick to move around in the x-y plane, for testing and basic board adjustment. Fiducial recognition and board alignment should be done by camera of course, but to give the software a hint where the board is the joystick allows to move to the approximate position.
I wanted to have the top metal platform made since my CNC is too small to mill it so I asked a company what it would cost to do it for me. The price was about €1.500, too expensive for now. Thats why currently I only mounted a 10mm thick and 200mm wide aluminum sheet I had. Enough to hold my small PCBs and a couple of smd parts.
Pick and Place Software
Since I already did write some code some years ago for a pick and place thing I had something to start from. Writing a gerber-file importer took me two days (I did it from scratch using the gerber file format specification from Ucamco). Martin adapted the old code and added the support for our motor drivers, our cameras as well as reading the pick and place csv files. So we had our complete software solution to load gerber and pick and place data and control our motors. Sweet!
For the vision I used OpenCV which I knew a bit. This was very straight forward, it took me under an hour to get a usable rectangle detection on an 0402 part with the rotation information. Still have to test how it will work in the real world. If I learnt anything from building machines than that in the production environment not a single thing will be as in your lab/test environment. I set the filters so that only the red light of my ringlight is captured. That way sunlight won’t screw up my detection. We have a glass roof in the office and no housing for the pick and place yet, so that was quite a good test for the relieability of the alignment algorithm.
Another cool thing when writing your own software is that it can be taylored to your specific process. Since I use GIT as my version control system (even for Altium) I want the pick and place machine to pull the project and load the PnP and gerber data. The PnP will of course have a wireless lan built in and a nice sweet touchscreen and nice GUI.
On what hardware should the software run? Raspberry Pi?
The PnP could be controlled with a Raspberry Pi with our RasPiComm+ and in the beginning I wanted to use this as a platform. But since our old software is built with WPF (windows presentation foundation, a very powerful GUI framework), porting the GUI would have been too cumbersome. The bussiness logic would be easy to port since Mono runs on the Raspberry Pi. But in terms of GUI there is simply no match for WPF out there. And I want to use a touchscreen to control the PnP. Easy to make custom UI controls with WPF, very time consuming with other technologies. And when it comes to the vision system a little ARM processor would definately slow things down. The USB 3.0 cameras are running at 15fps at full resolution, pumping 180MB/s into the OpenCV algorithms, easy for a modern CPU but not realistic for an ARM processor (there is of course no USB 3.0 support limining bandwith even further). I also considered the Xilinx Zynq since I have a ZC702 development board. I played around a little bit, attached a camera but after a day I realized thats not the way to go. In terms of development time nothing can beat C#. And CPUs are fast enough for what I needed, and power efficency is not a mayor concern here. To be up and running fast I also would have needed extremly expensive Xilinx IP which I would avoid at all costs (no pun intended). And then there is the problem with fast adoptions. I want to be able to tweak and parametrize the vision system algorithms for different part classes, easy and fast when done in pure software, hard and time-consuming when done in FPGA logic even with all the amazing IP Xilinx provides.
So in the end it is a boring mini-itx board with a 3.2GHz quad-core and 4GB of RAM. More than enough, and alltogether (including CPU, RAM, SSD and full-HD touchscreen) cheaper than the ZC702 development board alone.
The automatic feeders are not done yet, I just have some basic design approaches. Some say they are the hardest part. I don’t know yet, my focus was to get the PnP up and running. But if its too hard I will simply buy some. Currently I only have a simple aluminum board holding down the components with a polycarbonate sheet just for initial test run purposes. Simple but it works for now. I’m still a bit struggling with the resonance of the motors, at certain speeds small parts can flip. But on one hand the resonance can be minimized by tweaking the CL-parameters of the motor controller which I have not done yet , on the other hand I still want to design proper feeders and replace the stepper motors with BLDCs. So thats not a mayor concern for now.
My next step is tuning the pick and place process and do a video. And building feeders of course.
Here is a short speed/motor test with a joystick (no actual pick and place action yet, I’ll do that in a separate video):
UPDATE: First pick and place test with and without vision: