Flying camera

[andrew...@gmail.com]

Since upgrading to Duet with Mr Blacks CAN head I am actually enjoying my pick and place machine rather than being frustrated with it.

My last two annoyances are terrible Z-Home repeatability and the ELP camera resetting to "auto exposure" randomly.

The former I am hoping to fix with a hall effect end stop.

The later I am not sure about but I will try Linux to see if it is a windows problem.

After this I am going to try make some improvements to get higher placement speed.

The two big things are going to be feeders that don't require push/pull head movements and motion speed improvement.

The last thing as a challenge for myself - I was thinking of trying flying vison.

Has anyone managed this?  Am I kidding myself to think I can get it to work?

[Mike Menci]

See this; https://groups.google.com/g/openpnp/c/OHevwbHVdCA/m/qEPNKacrAgAJ

[andrew...@gmail.com]

I was planning on doing the flying frame capture in in hardware (synced to what the Duet3D is doing) not trying to get openPNP to sync camera capture to the motion of the Duet which it does not really know 100% what it is up too.

Like send a Mxx command to the Duet and then it runs through some canned motion and gives two frames to the main computer (nozzle 1 and nozzle 2)

I have no idea if openpnp could even deal with the image just turning up like that.

[simpl...@tuta.io]

>> See this; https://groups.google.com/g/openpnp/c/OHevwbHVdCA/m/qEPNKacrAgAJ

As far as I know, a “flying camera” is mounted on the head and performs component alignment immediately after the components have been picked up, on their way to placement. That would be by far the most time-efficient solution. But the commercial solutions are all quite complex, expensive, and difficult to build yourself, due to the required optics.

A “fly-by camera” on the other hand, is much easier to implement, but the key challenge here is the required precision of the trigger, a consistently short sensor trigger delay, and fast transmission of individual frames or image sections. But even then, you shouldn’t expect super-high precision (like with a classic bottom-vision system) - the deviations are often in the double-digit micrometer range with my homemade builds, but eventually it might work with 01005, who knows.

This has been discussed here many times, but the intended use of such camera systems completely misses the point of a typical OpenPnP solution, which is designed to be cost-effective, user-friendly, and suitable for prototypes and small-batch production.

[andrew...@gmail.com]

Flying.  Mounted on head.

I am designing a new head for my PNP to make it lighter for faster moves.  It is currently a bit tank like with way too much aluminium and stainless.

It looks like I can shave 100s of grams metal off and still move my solenoids to the head by changing to SMC_V114.  (SMC_V070 didn't seem worth the extra $50 for saving a further 5 grams each)

Whilst I was looking at redesigning the down camera to be coaxial similar to my current up-camera

https://groups.google.com/g/openpnp/c/8mhTAm9CNnE/m/NIfnip7xAAAJ

https://groups.google.com/g/openpnp/c/8mhTAm9CNnE/m/ZZZiL9zjAQAJ

I had an idea about making the the down camera be able to see the nozzles as well for at least small parts.

To make the coaxial lighting compact enough - the redesign was going to need folded optics anyways.  Why not add some motors and mirrors.

I think I can make a system that the camera can look down and with some smoke and mirrors the same camera can see the part on the nozzle.

Pick up part.  Move to top Z.  Send command to Duet.  Duet spins a motor around to get image.  OpenPNP processes the image.

The thing I don't know about is if OpenPNP can cope with mutiple virtual up cameras. with different image sizes.

N1 and N2 flying camera images will be optimized for small parts up to about 5mm.  Then the standard coaxial up camera would need to be used for your bigger parts like TQFPs

[andrew...@gmail.com]

I found some prior discussion

https://groups.google.com/g/openpnp/c/_rpBVp3Bkc8/m/SUUG58ZmBwAJ

[andrew...@gmail.com]

If anyone else was curious about how small SMC V115 solenoids are

Left - Standard relay recommended around here

Middle - Two Chinese standard small HF10 relays

Right - The Japanese SMC-V115 (actually Chinese clone I reckon even though listing claimed original)

The manifold block is just stupidly large compare to the HF10 solenoid.  I was planning on doing my own manifold to hold four solenoids (suck and blow) so I can save some more weight there.

[andrew...@gmail.com]

Sorry - did not think of weighing them before hitting send on last post

72g  - Standard STNC relay

53g -  2x HF10 with the manifold

11.3g - Just the HF10 Chinese unit

11.8g - SMC-V115 (probably clone)

So it should be pretty easy to have 4x HF10 and a custom manifold for less weight than a single STNC relay

On other parts of the head I have already saved several 100g in motors and unnecessary stainless steel.  So it is looking pretty good so far.

[SM]

I highly recommend the T10 valves, as I’ve been using this type for many years without any issues. For the vacuum manifold, I use 3/2-way valves (large flow), and for the blow manifold, I use 2/2-way valves from OST. The short hose lengths to the nozzle enable fast response times (approx. 5-7 ms).

However, the whole setup with the valves, motors and electronics on the head has one drawback: it all gets quite heavy, so you can’t do without ball screws with servo drives and a robust frame. On the other hand, if the whole machine is already as heavy as four washing machines, a few grams more or less don't really matter.

As for the flying vision, I really like the light weight solution from the comercial machine E4 (with CCD laser scanners on the head).

[andrew...@gmail.com]

Are the T10 you are speaking of the ones I show that the vendor called HF10 ?

Are the CCD scanner heads from "E4 machine" CCD with opposite side illumination and the nozzle rotates the part and analyses the shadow cast by the part to work out rotation (as Mr Elson has previously described)

I am also keen to investigate that idea and have purchased some TLS1401 for testing that idea as well as the larger CCD I have bought for the flying camera tests.

[Mike Menci]

It seams like this are 3way manifolds bat 2way selenoids - it should be 3way to work - 

[Mike Menci]

like here:

[simpl...@tuta.io]

>> Are the T10 you are speaking of the ones I show that the vendor called HF10 ?

I don't know. This (see images) is written on my vacuum (3/2) and blow (2/2) valves:

[SM]

Mike, the manifolds are always the same, just different caps.
Here's how i did it:

[andrew...@gmail.com]

Yes - those solenoids look like the same as the ones I have purchased called HF10.  The 10 in all of them is for 10mm I assume.

They are same width and bolt/port pattern as the SMC-V114.  Which is a bit larger than the SMC-V070 which is only 7mm wide but much more expensive.

The Chinese/Aliexpress manifold is quite excessive/heavy and is much heavier than the solenoids themselves.

I think with a custom fabricated brass manifold I should easily be able to get my overall head weight lower.  If I just got my current head and cut to length the thick walled seamless stainless tubes on my C axis I would save the weight of four of these solenoids.

Even a custom machined ally manifold could be 1/4 the weight of two of the chinese ones back to back for suck/blow.  And milling from billet is much easier these days than fabrication from parts.

BACK to the "E4 CCD scanner head"  was it the idea of using a CCD to look at the shadow of the chip resistors and capacitors as they are rotated around C

If so - has anyone built that yet for an OpenPNP machine?  It seems like it should be light to build and computationally trivial.

I have purchased parts for trying a flying head 3 different ways and the chip shadow method will be the easiest if it can work.

[Mike Menci]

Well done! 

Poslano iz iPhon-a

Dne 17. apr. 2026 ob 12:38 je oseba 'SM' via OpenPnP <ope...@googlegroups.com> zapisala:

Mike, the manifolds are always the same, just different caps.

Here's how i did it:

--
You received this message because you are subscribed to the Google Groups "OpenPnP" group.
To unsubscribe from this group and stop receiving emails from it, send an email to openpnp+u...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/openpnp/ace1de27-fddf-4d54-9b03-28831c4334d2n%40googlegroups.com.

<pneu.PNG>

[andrew...@gmail.com]

I'd say from the timestamp in this video about a "Tonstrol E4" 

https://youtu.be/fBYNy7gVhmQ?t=161

Confirms you were talking about a linear CCD looking at the width of shadows (Like Jon Elson mentioned many years ago)

It seems like the easiest of my three ideas I want to try implement I just am not sure if it works for SOT23 like parts or if it is only useful for chip resistors/capacitors.

I think the next easiest idea of my three in terms of mechanical complexity is the linear CCD that sweeps past the components.  That would work for SOT23.

For anything larger than SOT23 or prob DFN10 I would be falling back to a standard coaxial up camera.

[simpl...@tuta.io]

>> I think with a custom fabricated brass manifold I should easily be able to get my overall head weight lower

Brass is quite heavy compared to alu.
A valve weighs about 13.4 grams and a 4-way manifold weighs about 56 grams.

>> I think the next easiest idea of my three in terms of mechanical complexity is the linear CCD that sweeps past the components

I haven't looked into this very interesting topic yet, but I think you need two sensors offset by 90 degrees.

[Mike Menci]

I use single as well - single nozzle -light weigth ... 

[andrew...@gmail.com]

Brass is denser than aluminium.  However fabricating means you can use much less materials than CNC milling which should make up for the 2.5 times heavier material.  Plus each port is already a brass tube that the low pressure/vacuum lines can just push fit on so no need for threaded pneumatic couplers.

I will CNC ally ones or 3d print some to start with just because it is much easier to do.  But if i need to save that last 10 or 20 grams I have the brass option.

The extruded and machined Chinese ones are not trying to be light or compact.

>>> I think the next easiest idea of my three in terms of mechanical complexity is the linear CCD that sweeps past the components

>I haven't looked into this very interesting topic yet, but I think you need two sensors offset by 90 degrees.

Just imagine this

https://elm-chan.org/works/lcam/report.html

mounted on an MGN9 linear rail so the linear CCD can move past the nozzle - rather than the linear CCD staying still on a tripod and the train to Akihabara moving past.

From a mechanical perspective this is easier than option 3 - trying to use an area sensor camera and mirror that flips up.  If your mirror does not flip to correct angle then you have problems.

Even when the linear CCD on a linear rail gets weight reduced from a big old heavy MGN9 to something else - it is still not going to be as wobbly as an area camera trying to not shake.

[SM]

>> Just imagine this

As I said, I’ve never looked into the holy grail of component alignment, but if the nozzle pulls the chip upward against the (fixed) CCD sensors, the component height - or more precisely, the exact lower edge of the component - can be determined. And since rotational movements also take time, two sensors offset by 90 degrees (with sufficient resolution) are likely a good idea, since you only need to rotate the chip 90 degrees to obtain usable silhouette data.

By the way, I’ve also built a head using very small, lightweight MGN5s - but they had too much play. Currently, I'm using affordable MGN7s with long carriages on the Z-axes.

[simpl...@tuta.io]

I think only one CCD sensor should be enough. 

The component is first rotated a few degrees back and forth, then rotated 90 degrees, and rotated a few degrees back and forth again to determine its alignment based on its silhouette.

[andrew...@gmail.com]

Option 1 there also seems like it would be perfect for nozzle runout as well.

Maybe even if option 2 works well it would be worth the extra grams to keep option 1 for nozzle runout and chip resistors/capacitors.

[simpl...@tuta.io]

I would start with this sensor:

https://www.mouser.ca/c/?q=TCD1254GFG%288Z%2CAA2%29

and a 5-10mW laser diode

[andrew...@gmail.com]

Is there a reason you chose that particular sensor?

One of the ones I have in the mail heading to me is the TCD1103 - which is 8.2mm wide.  The TCD1254 there is 13.1mm wide.

I only think the option 1 "shadow on the CCD" type component alignment works for chip R and C so 8mm should be wide enough yeah?

I also got some TSL1401 which are only 128 pixels across - but I wanted to see if sub pixel intensity could get decent angular resolution.

The other one I thought was interesting for option 1 was the EPC901 which is 8mm and 1024 pixels.  it is a nice flip chip package that would be easy to integrate.

For option 2 (line scan camera) I got an obsolete TCD132D just because it is easy to work with in experiments.  I think the optimum CCD on the market for that task is the Hamamamatsu one.

Option 1 the non square pixels is good.

Option 2 would prefer square[ish] pixel sensors.

[simpl...@tuta.io]

>> Is there a reason you chose that particular sensor?

Yeah, it seems to work well in the machines mentioned above.

I would set the distance between the sensor and the laser diode (without focus lens) as far as possible to ensure good illumination.

You can get the ccd and a laser in china for a few bucks and interface it with a blue pill etc.

[Michael Anton]

The laser and CCD method sounds a lot like what CyberOptics does with their sensors: https://www.ebay.com/itm/267572747667.  These have been used for component centering for decades on many commercial machines.  The old ECM93 that I used to run has one of these Cyberoptic sensors, and one of the machines I bought also has one.  The one I bought I stripped down for parts, but kept the CyberOptics sensor and electronics.  IIRC the electronics spit out a serial data stream that specified the component width and center in the current view.  Then the part is rotated to find the minimum, maximum, and center.

[andrew...@gmail.com]

From the moment I read the Jon Elson post about some ancient commercial machine using the idea I thought it sounded like a fairly easy way to get "flying" rotation correction for at least small rectangular parts.

It is going to be a lot easier than the full flying camera which is going to require moving mirrors with all the problems of shaking and alignment.

The line scan camera with mirror is still the option I think looks the coolest - and I will try all the options.

My TCD132D arrived in the post today.  Still waiting on the TCD1103 and the TSL1401.  Have just ordered some laser diodes with line generation optics but first test are going to be with LEDs at a very long distance