Adding Vision to Delta Robot and Other Upgrades 

I may end up going that direction eventually but it's fine for my applications right now. I have been eyeing the OpenMV Cam H7 which comes out in march.

Can you explain me that with the raspberry please?

Engipreneur openCV is a computer vision library for python that is actually used in commercial applications and very easy to learn just youtube that you'll find quite a few of those vids explaining different element of CV

i did some color classification with openCV, but be aware that a lot of camera's have inbuild auto white balance (awb). you can set a lot of camera properties with openCV, but not awb for some obscure reason. You can disable it manually with CV_CAP_PROP_SETTINGS, but most camera's dont remember their settings after power loss, so you have to do that every time. There are ways to work around it (i found a way using Direct Show, which belongs to the Software Development Kit of Windows). Maybe also solutions for raspberry exists? But awb can be a real nightmare.

​@@maatwerkengineering3398 The raspberry pi has APIs for getting/setting coefficients for AWB and allows auto AWB to be turned off so you can get coefficients that work and stick with them. I was able to do black and white image processing for aruco marker detection at 25fps getting half sensor resolution(1640x1232) 4:0:0 chroma subsampled frames from the GPU (1640x1232 black and white + 820x616 red/blue components). The GPU prefers to output YUV formatted frames since it's optimized for video encoding and video is typically chroma subsampled meaning you get half the color resolution in each dimension. It can output RGB formatted frames but that slows the capture framerate somewhat. Color detection in the YUV space is completely practical though. Overall it's an OK computer vision platform that beats anything in it's price class.

Cool, I've never used those. Thanks!

Thanks! Yeah, I should also probably paint the table matte black or gray as well because it definitely picks up the reflections on the aluminum.

I don't need to plug the camera into the computer, that was just to show everyone what the camera sees. The camera stores all the settings on board and communicates with the ESP32 via SPI. I will eventually get involved in OpenCV 4.0 using either a raspberry pi or just use my spare laptop. This is my first attempt at anything involving machine vision so once I get comfortable, I'll move on to more advanced features of OpenCV like facial recognition and neural networks.

@@ProtoG42 Well good luck.

It depends several factors like how fast you want to go and what range of motion you need. Backlash is only an issue if you are placing very small objects and it is easily compensated for with the use of encoders for absolute position. This is the reason I placed the encoders on the output shaft of the gearbox and not on the back of the stepper motor. If the encoders were on the back of the stepper motors, they would not be able to measure the backlash in the gearbox. Without a gear box or other belt driven gear reduction, a delta robot like this with linkages will easily stall out if you try to move too fast. I'm using a 5:1 gear ratio to and I'm able to travel at a much faster speed than if I was not using the gear box. Also, it adds a bit of holding torque without the motors having power applied. Without the gearboxes, the machine will just faint when power is removed. Microstepping is required if you want any respectable amount of precision. The torque of a stepper motor is drastically reduced when microstepping and when driven at higher speeds. At 1/8th microsteps, you're already at around 19% of the original rated torque. At 1/16th microsteps, the motor now only has around 10% of the original torque rating. The power of these motors gives me much higher torque at high speeds and higher resolution compared to smaller stepper motors. Stepper motors are not usually subjected to direct loads like this as they are most commonly used where they have a mechanical advantage, like turning a lead screw on a cnc machine or a pulley on a 3d printer. The farther away from the shaft that the load is applied, the more difficult it is for the motor to move quickly.

@@ProtoG42 thanks. so I will use a 5:1 reduction with belt reduction. because is expensive a harmonic drive. I think belt reduction has zero backlash with htd belts. thanks man

Thanks for the tip! I have looked into using a linear motor with a syringe but I never thought to use an air piston. I have plenty of air cylinders laying around I could use.

that may be the way to go in the future, but for now, I only need a snapshot of the table before it sorts the balls so this setup is fine for that application. Real time tracking would definitely need multiple cameras.

well they do look like fruits and also they are glossy implying that they are full of sugar

Multifold cup handles those fine, designed for rough irregular surfaces. Does have a longer travel though, so less precise positioning, but generally used to handle paper in a printing press for this reason of sealing well.

@@SeanBZA Exactly! These Schmalz cups are great and make it surprising difficult to remove the golf ball when the vacuum pump is on.

Nice project, just some advice. The shinny silver aluminum ball support tray is no good, make it flat black and your camera will work much better! Your video would make more sense if the golf balls were in a random color pattern on the tray, in this video they are already sorted.

​@@glennmura I needed to place all of the same colors together to make sure the camera was identifying individual balls instead of one larger object. As I mentioned in a previous comment, I will probably paint the table matte black or gray because it definitely picks up the reflections on the aluminum. The robot will end up sorting the balls with different common sorting algorithms. I think it will be a interesting way to visually see the differences between algorithms like bubble sort and insertion sort.