(Oops saw other comment linked, but I’ll leave this comment in case summary is helpful)
Check out the SO100 arm, being supported by Huggingface and others. Only 4 axis, but cheap AF (<$250/arm) and using ML to make it more capable than 4 axis would seem. Also using identical arms for mirror teleportation/training.
4 is still though fundamentally limited: you need a minimum six DOF to position all three translational and rotational freedoms. The human arm has 7, and I think a case can be made the smart software would reduce the problems from having only exactly as many as required. ... but not eliminate: the extra degree of freedom lets you "get out of your own way" when moving objects that are not zero size trough multiple positions :)
Perhaps 4 is enough for any specific application, but then again perhaps 3 is or 2. :P
I've never programmed a robot arm, but I've spent a fair amount of time using a seven axis faro arm (a coordinate measuring device, sort of the opposite of a robot arm) and it certainly takes some practice to avoid "cant move there from here without reorienting everything", it's easy to take for granted what our brains do automatically for us. :)
The point is that you can achieve a lot without being able to access every position and the human arm having a lot of degrees of freedom doesn't necessarily mean it's more capable than a different arm with fewer.
Sure but why not, say, an XYZ or XYZT gantry instead?
There are a lot of downsides in 'arms' -- stacking all the axis hurts their speed, stiffnes, accuracy, increase complexity (e.g. in pathing, limiting speed, restricting to safe areas), limited reach.
I've always thought of them primarily as beneficial for flexibility at considerable cost, but if it's not 6dof then the flexibility isn't so great, then why not some other geometry?
in simpler term: an object has position and rotation, and we're in 3D, so we need minimum 6 linearly independent parameters to be able to both point at a direction and be at a location at the same time.
Drone and helicopters has 4, and they are able to control max 4 of 6 parameters. Usually 3 positional + 1 rotational, and the rotatinal axis go first.
It uses cheap hobby servos which is, to put it mildly, not the same as steppers. You will never get 0.2mm precision with them. They will were out quickly and even quicker under load.
BTW, looks like it doesn't have closed loop control. "0.2mm repetability" (should be repeAtability) is only 'under certain conditions', no load.
You'd be surprised what you can do with servos modified with encoders, setup with an industrial grade cascading control loop. First demo video below demonstrates threading a mechanical pencil lead in and out through the writing tip. Second video demonstrates a 187gram weight at the end of a 470mm long rod. Third video shows that the modification is now a quick fitting of some 3D printed parts to the servo motor axle and servo housing.
It has closed loop control, these [0] are a bit more proper serial servos with digital absolute magnetic encoders. Each one can be sent positional, velocity and even acceleration targets. I doubt anyone is looking for 0.2mm precision for $300 total. Or even 1mm precision.
The PAROL6 BOM doesn't provide any cost estimate, but it looks quite a bit more expensive.
I really never understood why hobby servos are so shitty. Their protocol uses only a very narrow range of PWM values and not the full 0-100% scale and that alone completely wrecks their precision beyond what their motors can achieve.
Also they do measure position to achieve their feedback, might as well just output that on a 4th wire.
It's because the default "analog output" PWM mode of a microcontroller will only give a rough approximation of the signal that the servo actually requires. For a servo, the duty cycle is (almost) irrelevant, the 0-100% scale has no meaning here. What matters is the actual length of the control pulses in milliseconds - the gaps between them can be arbitrarily long within a certain range.
If you think about it, it actually makes a tiny bit of sense. First, it is failsafe: Breaking the control line or shorting it to ground will not move the servo to 0%, shorting it to signal level will not move it to 100% - it just doesn't move at all and stops applying force. Any sentient being within the movement range will definitely prefer it that way instead of random movements. Second, it can actually be pretty precise: The driver circuit can be completely analog, it doesn't have to be limited by arbitrary digital quantization steps. All it needs to do is check if the current encoder value is above or below the target and apply power to the motor accordingly.
They've been around for a really long time. I suspect that back in the analog radio days of yore, the control pulses transmitted by your Futaba were recieved as a pulse train - 6 channel radio, six pulses and a big enough gap between frames to reset the index. At the RX all you have to do is feed those pulses through to successive servos one after the other, which is easy and cheap to do with basic logic ICs.
So when you use hobby servos for robots, you're taking a low precision actuator meant to make a flap or throttle or other control surface go relatively up, down, in, or out, that was designed in like, the 70s, and asking it to do modern robotics stuff
My guess would be they use variable resistors for encoders. I haven't done servos for a while, but the industrial stuff all used optical encoders with Gray code patterns.
There are bit more expensive servos with proper encoding. The thing is hobby airplanes don't need high precision as pilot controls looking at the whole plane movement. Now that hobby is killed by regulations. Robotics is coming..
Check out the SO100 arm, being supported by Huggingface and others. Only 4 axis, but cheap AF (<$250/arm) and using ML to make it more capable than 4 axis would seem. Also using identical arms for mirror teleportation/training.
https://github.com/TheRobotStudio/SO-ARM100