It's not yet clear to me how one programs a robot dog with a "moonwalk" envelope while also ensuring it doesn't fall over, but hey, who am I to discourage hating on something that genuinely strikes fear into my deepest bone marrows?
It doesn't seem feasible to be programming these to dance by programming all individual actuator movements - there's got to be a separation of goal and execution, and somewhere in the execution the foot placement and force is tweaked to keep balance. Inverted pendulum bots do the same thing on easy mode with wheels. A simple inverted pendulum bot is something easy to do with a PID controller and much harder to control with scarequotes AI and machine learning. I'm sure some high schoolers have done it by now. I found this interview with Boston Dynamics CEO from 2019, which is mostly just fluff but it seems like they're going with "AI" for the low level reflex-type movements. Personally I have to say the dog-type robots are farther along the uncanny valley - the humanoids have a "chair-shaped-ass" leg and hip posture that isn't threatening and it's hard to dance with such an underdeveloped booty.“I break AI down into two parts. I call it athletic AI and scholarly AI. We’re kind of experts at athletic AI
I suspect it's simpler than that. I'm rebuilding a 3-axis mill. It'll be a 5-axis mill with ATC when I'm done, so six axes. 30 years ago those were just AC servos. Now? Well, now you you need to start this video at 3:57. Back when I was a strapping young lad the cheaters put something called a "Sabine feedback destroyer" inline with their conference systems. What it did was analyze the signal for what mathematically looked like runaway feedback (the "ringing" you hear sound systems do whenever someone in Hollywood wants to create tension at a High School dance). It would then mathematically apply a filter to the center frequency of that resonance and the feedback goes away. That was over 26 years ago. The difference between electrical and mechanical oscillation is... nonexistent from a programming standpoint. You use the same math to model them. It's all Diff EQ, and if you sample the output and the feedback often enough you can balance a dozen angels on the head of a pin. I'm running Sigma 7s, which run at 3100 Hz. So... every 0.3 microseconds, the servos compare "where am I" to "where am I supposed to be" and adjust accordingly. Turns out, Boston Dynamics uses Yaskawa, too. After that it's all degrees of freedom. Atlas had 28, whatever they're doing now prolly has more. Makes for some yucky math - and I'll bet their "AI" is nothing more than "train the simulator to goal-seek for stability with a few dozen degrees of freedom" and then they just overlay macros on top of that. The happy dances, then, are macro overlays on an armature that fundamentally keeps itself in pose through feedback loops. Here's a guess - a dog's hip joint is pretty easy from a mechanical standpoint. It's got one degree of freedom. That's a planetary, or a harmonic drive. Doesn't matter, I own a half dozen of them. A human's hip joint, however, is a ball joint. Closest you can get to that is a hollow-shaft harmonic drive with a right-angle harmonic drive through the middle. You can get the motion you need that way but it's gonna look like this: https://www.fanuc.eu/~/media/corporate/products/robots/m20/m-20id-12l/m20id12l.png I've been looking at this stuff a lot lately and I gotta say - I dunno how I'd do "hip joint" using mechanicals.but it seems like they're going with "AI" for the low level reflex-type movements.
Personally I have to say the dog-type robots are farther along the uncanny valley - the humanoids have a "chair-shaped-ass" leg and hip posture that isn't threatening and it's hard to dance with such an underdeveloped booty.
The power, efficiency and speed of shape memory alloy - which is what all these projects rely on - does not lend itself well to power transmission. I ended up getting a super-duper tour of the Microsoft design lab and they had some stuff out. I looked at one and went "holy shit - is that nitinol?" and the engineer tried real hard to keep the shit-eating grin off his face." "Yeah, that's nitinol." "You guys are prototyping nitinol actuation?" "We're in production on nitinol actuation." That was 2018. If you look up "microsoft surface detach" on Youtube you will find a couple dozen videos about what to do when it doesn't work. That's how far along we are with "soft robotics:" bleeding-edge tweaker product designers sneaking it into halo projects and having it not work. Pneumatic systems work great, so long as you understand their limitations. Air is compressible. That makes everything tricky. You also have a limit of "hard vacuum" which is why any schlub can pull up a syringe with a cap on it, unless that syringe is like an inch and a half in diameter. 14.7 psi. That's the speed of light. You wanna pick up a 100lb load using pneumatics in retraction? You need a tube more than an inch and a half in diameter. And since you're mostly doing this stuff around joints you're talking torque and 100 in-lbs is minor leagues. Hydraulic systems? Also work great. Also aren't fast. A hydraulic piston needs a hydraulic pump somewhere and in the end, it's nothing more than force conversion. The whole of the animal kingdom operates on force through retraction and to date, the only material we have that works like that is SMA: Kid's shows were freaky once. Nickelodeon had a show, that wasn't any of these, about a kid who ends up befriending a juvenile alien that was of course an invisible cloud because it was iTV and cheap. The kid is showing off his family's new car to this space alien who laughs and points because humans were so obsessed with wheels. Some screenwriter in 1970s England was making a commentary on how pretty much everything the human race has ever done is some form of rotary motion and he wasn't wrong. Obviously it was formative for me. There are positioning systems that don't rely on rotary motion. Linear servos are definitely a thing. They're even dumbly accurate. But they aren't muscles, and won't be any time soon.The challenges in designing SMA applications are to overcome their limitations, which include a relatively small usable strain, low actuation frequency, low controllability, low accuracy and low energy efficiency.