The article seem to be referring to the problem of manipulating a physical Rubik's cube. In that case, the bottleneck is the mechanical interaction, not the computation of the necessary steps, right? The real achievement here is the speed of robotic manipulation?
It's kind of misleading to phrase it as the chip solving it, when the focus is on the actuators.
They didn't even improve actuation. Instead, they appear to have used a specially-constructed cube (no doubt designed to be suited for their machine) and literally bolted their actuators to the center caps. I'd like to see them do this with an unmodified vanilla cube like the other cube-solving machines I've seen.
If you want impressive actuators, check out modern pick-and-place machines [1].
Bolting the caps to the actuators was the bigger trick anyway. Compared to cages or grippers, dramatically reduced inertia and much stiffer, so less damping required and it's easier to dump energy back into the actuators. I'm half-tempted to call that "cheating", but I don't know what the actual rules are - I doubt that this robot can "put down" or "pick up" the cube in less than a minute or two. The older version had the cube actually soldered in.
The chip is targeting ADAS (Advanced Driver Assistance Systems) applications, and I would assume that the hard part is to run vision algorithms (that is responsible for reading the cube and/or visual servo) on an automotive MCU, which almost always prioritizes safety and standard compliance at the cost of performance and has weird custom architecture (PowerPC in the auto world is already "mainstream"), at a really high frame rate.
You don't need "vision algorithms". Just average the color of a couple of pixels per square. The cube is always in the same place, no need to do anything fancy.
That'll be because that's what the press release said. Any seasoned journalist would have re-nosed this story to lead on the fact that it is a machine that manipulates a physical cube very quickly and poor old Infineon would not even have got a mention.
It wouldn't be doing any image processing during the solving phase. In fact, it's probably not doing any full stop, and just tracks the state of the cube (which starts in a known state) via dead reckoning.
They use stepper motors - same hardware that allows a budget 3D printer archive sub-mm movement without any feedback loops.
Pretty sure they don't need to look at the cube again after scanning the starting position.
Fast stepper motor control was solved a long time ago, and takes little computer power. Watch this daisy wheel printer, a technology from the 1970s.[1] The print wheel is driven by a stepper motor, and for each character, it must be spun to the correct position, brought to a precision stop, and held while the hammer hits the selected letter tine.
Daisy wheel printers were the first mass-market device with multiple powered mechanical moving parts coordinated entirely in software. The four actuators (daisy wheel stepper, hammer actuator, carriage stepper, and platen stepper) were connected only by electronics and software. Invented by David S. Lee.
Looks like a MoYu Weilong stickerless cube. I think it can handle that speed without modifications.
Any 'legal' rotation is pretty easy on the insides, mechanically. You can adjust the tension and add some lube. It doesn't seem to do any corner-cutting in the slo-mo part of the video.
I guess this mimics human competition settings: you're allowed to examine the whole cube beforehand. Figuring out any valid solution is really trivial, esp. on a computer, it's just not hard.
"The microcontroller AURIX™, one of the world’s most powerful minicomputers, also contributed to the record-breaking effort: It is one of the essential elements that enable autonomous driving."
I do not think the term 'minicomputer' was intended to mean... a microcontroller.
>So it doesn't have to do the 'minimum' moves as they say, might waste time working it out.
Maybe they're counting moving the top and bottom at the same time as two separate moves? I don't know why it wouldn't just be considered moving the middle but that might be what they're talking about. They mentioned that it doesn't do it in the minimal amount of moves on the basis that it would take longer so I'm assuming that's it.
As far as the computational time needed to do so, it's absolutely trivial. This thing is built around a single microcontroller and it can handle 6 stepper motors, solving the cube, vision processing, and state tracking all by itself. The only thing impressive about this is how fast the cube is being manipulated but even that might not be too impressive considering the fact that they just attached stepper motors directly to the center piece of each face as opposed to using an unmodified Rubik's cube.
Even then, this seems like a poor choice of demonstration of visual recognition capabilities since you only need to read the initial state and then adjust it when you make a move. They've used stepper motors which normally don't need feedback too, do movements can probably be done open loop.
Perhaps they do only use visual feedback for determining the state, which would be a better demo, but I doubt anyone would be able know from that article!
It doesn't. I can be done quickly on any conventional computer: https://en.wikipedia.org/wiki/Optimal_solutions_for_Rubik%27...
The time is still impressive, but this doesn't showcase any real improvements to any technology.