I spent some time on legged locomotion back in the 1990s. It was clear then that you wanted torque control, and I did some work on the theory for that, trying to solve it from first principles, not machine learning. Got some nice theory and a patent out. But the parts just weren't there to build such things. As the article points out, the key to this is motor back-drivability. The final drive has to survive shock loads, and it has to dump forces into the motor, where the magnetic fields can take it. As I've quoted before, "you cannot strip the teeth of a magnetic field", a comment from early General Electric locomotive sales. (Locomotives are Diesel-electric, not Diesel with a clutch and shifting gearbox, because the clutch required is huge. Yes, it's been tried.)
That's something few areas of engineering cared about, with the exception of aircraft flight control systems with mechanical backup.
Pneumatic actuators looked promising, but proportional dynamic valves were big, heavy, and about $1000 each. Linear motors (not ball screws) looked like the coming thing back then, as 10:1 power/weight ratio had been achieved.
But that technology never got much further, and Aura, the biggest player, collapsed in a financial scandal. Series elastic actuators were (and still are) a race between the spring compressing and the ball screw motor starting up. Hydraulics were too clunky; Boston Dynamics built a 400 pound mule, but the Diesel power pack never worked.
Direct drive pancake motors were used by some SCARA industrial robots, but those were too big for leg joints.
I thought someone would crack the direct drive problem eventually, but nobody ever did. We're still stuck with some gear reduction.
Some of the exotic ideas for muscles mentioned in this article go back that far. The McKinney muscle is old, and not too useful. There was some interest in electrorheological fluids, fluids whose mechanical properties change when an electric field is applied. That didn't become useful either. Shape-memory alloys were a dead end; liquid cooling can overcome the slowness problem, but not the inefficiency problem. Everybody went back to good old electric motors, although they became 3-phase AC instead of DC. It helped that the drone industry made 3-phase motors and their controllers small, cheap, and powerful.
Academic robotics groups were tiny. MIT and Stanford had less than a dozen people each.
Progress required hundreds of millions of dollars for all that custom engineering and R&D. The level of effort just wasn't there. Nor would throwing money at the problem prior to machine learning have led to useful products.
It's impressive what's been accomplished in the last five years. It took a lot of money.
Silly question maybe, but didn’t Boston Dynamics have videos of bipedal robots doing acrobatics / running ~7/8 years ago? Kinda looked like they “solved” locomotion then
Their approach required pre-computation and simulation before execution. If you watch their videos carefully, you can see the advance planning work on some of the screens.
Some of it was pre-computed. The middle layer, if you like. The Boston Dynamics group had walking gait of a sort nailed in the 80's; the trotting-on-the-spot that BigDog did was essentially a continuation of those mechanics and that's all based on a conceptually simple balance problem which is intrinsically reactive and not pre-planned. So that's what was going on at the lowest level.
At the top level you have the actual environment, with those meme videos of the robot trotting through a car park, getting kicked off balance, and recovering. The whole point of those tests was to demonstrate how robust their tech was to non-precomputed disturbances.
And between the two you've got the direction and planning layer, telling the robot to go from A to B with some set of suitably convoluted parameters that nobody but the operators would have understood. That planning layer might do all sorts of pre-computation and simulation but it needs to do it in the context of a noisy and possibly adversarial environment. That's equally true for Atlas as much as it was for BigDog, even when there's nobody actually kicking it. What I suspect the precompute and simulation is doing at that layer is a) checking for physical viability of the requested route, and b) parameter tuning in response to sensor readings over a number of runs. Not telling the robot the exact sequence of motions. But I'm nowhere near those teams (oh, I wish) to comment on whether that's true - maybe someone else round here is.
Right. Early thinking on this was zero moment point (ZMP) control. The zero moment point is where, if you land from a stride there, you continue going in the same direction at the same speed. You can speed up by displacing the landing point backwards a bit, and slow down by displacing it further, and turn by displacing it left or right.
When a foot or feet are in contact with the ground, the goal is to stabilize the posture. When in flight, the goal is to hit the zero moment point.
Reactive controllers do that. The planner picks the next landing point.
ZMP robots tend to start moving by running in place, then slowly transition to fast forward motion. You see a lot of robot videos like that.
The next step up is to control takeoff. Bend knees, fall forward, launch. Launch angle is a planning problem. Speed and direction changes become much faster. Basic gymnastics are almost within reach. The planner controls launch angle and landing point, and the lower level reactive controls do the millisecond-level control.
It's possible to do all that from first principles of classical dynamics. That level of performance was reached before machine learning. That's what you're looking at in older BD videos.
I can understand pre-computation making the “software” problem of locomotion easier, but how does it help with the hardware problems laid out in the article, ie repeated very high load over a very short amount of time?
BD used hydraulics for a long time. Works, but inefficient. You have to carry the actuators, the tank, probably a hydraulic accumulator, the pump, valves, and the power source for the pump. That's why BD's machines were so big. Someone at Google said "We need to have a conversation about hydraulics", and the dog robot in 2019 was the first all-electric machine.
BD was under business pressures, and a computerized automaton doing baked ninja back-flips with servos is more impressive than inexpensive FK/IK demos dead-lifting 1000lbs. Google broke that company with their opinions.
Even if private labs have a viable platform solution, people won't care unless they can clone it for free. Not a lot of incentive for design change, but building Kryten 2X4B-523P would be hilarious. =3
overclocking a CPU might make it seem that you solved something and gotten better performance, but sooner or later it breaks down, as I read the article I believe that pre-computation essentially allows you to "overclock" the hardware, and make it seem that you have solved the problem of locomotion when what you have actually done is made something that looks impressive for a very much shorter of time than is usually used to calculate what the hardware can bear.
Nobody has a problem with companies using AI to edit articles, create images.
But when even the writer is an AI persona, the trust factor gets destroyed.
You evaluate the merits of the content? From a high-level systems pov, the article is largely correct, even if some details might be missed / simplified
How do you evaluate the merits of a graph of information? I usually read articles like this to learn something, not to grade someone else’s assignment.
As a lay person, I have limited knowledge of this field, yet am extremely interested.
Unfortunately AI gen content is being used widely to spread spurious information/fake news etc. So my knee jerk reaction to AI gen content is - "this is going to be fake".
If the information you are trying to convey is true, and is technical/objective in nature, then why shy away from associating your identify to it?
Right. I'd noticed that the first image, of the harmonic drive, was wrong. The flexible ring gear can't have a disc closely attached; that would kill the flexibility that makes the whole thing go. In some designs, the flexible ring gear is just a ring. The inside of the ring has to be smooth. The elliptical plug is way too small; it has to fit the ring gear.
Harmonic drives are confusing enough. It's amazing that they back-drive at all.
Anything wrong in the text? I'm out of date in this area, and I was reading this as a nice catch-up. What else did they get wrong?
Even the video is generated. The whole thing is just slop upon slop, I'm amazed that it got to the top of the front page here. I suppose it's a genuinely impressive amount of fakery all integrated together.
This is AI slop and the article contains some of the worst illustrations I have ever seen. Most do not make any sense mechanically. Here are the worst ones:
- The "orbiting threaded rollers" in figure 6 are not meshing with anything (not that they could, since they are orientated in the wrong direction).
- The ball of the ball screw in figure 7 deforms the screw and the roller screw "meshes" with a flat surface.
- The guy on the pogo stick in figure 14 is jumping himself rather than putting his feet on the stands of the pogo stick.
- In figure 16, a key penetrates the elastomer skin of the optical tactile sensor, destroying it.
Is it? Now I feel bad for having posted it. I mean, I know some stuff but not enough to judge over all the content in this article. It's unfortunate but from glancing over I thought it was a comprehensive and useful resource. I guess I will just give up on the Internet or something like that.
« Rotary Actuators (The "Reflected Inertia" Trap) », « Quasi-Direct Drive (QDD) — The "Cheetah" Approach «
The pattern ‘something — The « metaphor » <qualifier> ‘ screams Gemini. Gemini seem completely unable to generate a section title that doesn’t follow this annoying pattern.
You just need some AI literacy, this document is very plainly AI generated. You don’t need to be an expert in the content to see that. First time you see an em dash your AI sense should be tingling. Then you see random bolded words alarm bells should be going off. Next you notice 70% of the headlines follow the pattern “The X: Y” and that seals it.
Actually, most videos seem to be real, run-of-the-mill showcases of various actuators the company sells. However, the owner seems to have added quite stupid AI-generated preview images lately.
There are a few of these being sold as products: AGIBOT has some models like that (eg https://www.agibot.com/products/A2_W). One argument that could be made for legged robots is that these wheeled ones can only work in wheelchair-accessible spaces. Legged robots can also balance themselves dynamically: a wheeled robot may tip over if anything violates its static balance, eg. carrying a load high up and going through a steep incline, though I guess having the torso be tiltable as in https://www.agibot.com/products/G2 addresses that.
Legged robots overall have more implementation complexity, spend energy just to idle standing up, but can go over much more varied terrain provided the controller is good enough. There are ways to adapt wheeled bases to different terrains (eg. larger wheels, whegs, RHex, rocker-bogies) but we know how to use legs to locomote over many terrains from personal experience, while the perfect wheeled/non-legged locomotion system perhaps remains to be designed.
There's also the way robotics is going toward data-driven methods, which in some forms (ie. imitation learning) require human teleoperation data. Here having the robot mimic the human form makes the mapping from human joints to robot joints easier (compared to other morphologies where you'd need to figure out how to best approximate a human motion with the joints/joint limits your robot has, though this is not impossible).
Presumably for outdoors or home deployment. The world is designed for bipedal locomotion, and human bipedal locomotion is designed for the world.
But yes, for a factory or commercial environment it doesn't make too much sense. It would be cheaper to adapt the environment, and many commercial environments are already designed to be accessible for wheelchair users anyway.
> Besides that, our entire technology is based on the human form. An automobile, for instance, has its controls so made as to be grasped and manipulated most easily by human hands and feet of a certain size and shape, attached to the body by limbs of a certain length and joints of a certain type. Even such simple objects as chairs and tables or knives and forks are designed to meet the requirements of human measurements and manner of working. It is easier to have robots imitate the human shape than to redesign radically the very philosophy of our tools.
It seems that he wrote that in a book published in 1953, but it's weird, I find, that he was imagining a robot driving a car. I would have thought he would have imagined that cars would become robots well before there would be humanoid robots wanting to drive them. So by the time you have a humanoid robot wanting to drive a car it's just one robot talking to another robot, electronically. And knives and forks are for eating, which humanoid robots presumably don't need to do, and is it likely that humanoid robots will need chairs in the same way that humans do? Altogether, a bad set of examples, I find. Perhaps the thesis would be more convincing with some better examples.
I need to get on to one of the prediction markets.
I’m willing to bet we’ll have a humanoid robot that can drive a car before we have level 5 autonomous vehicles.
And by can drive a car I mean a general purpose humanoid robot that can do basic household chores like move the car and wash it with a foaming brush and a hose.
I don’t mean the robot will be capable of sitting in any regular car and doing level 5 autonomous driving.
So you’re wanting to bet that we’ll get humanoid robots capable of driving a dumb car at L4 before we get cars capable of L5? When we have no humanoid robots driving cars, and many L4 cars driving around? I’ll take that!
But he was wrong already at the time. The world is made with humans in mind now. As soon as robots that are not human-like become useful enough, the world changes and adapts to them.
Just like the world had no asphalt and curbs before the car. Or just how heavily automated manufacture lines in factories have nothing that looks like a human can interact with it except a few buttoms here and there.
1. Asimov wrote that because he needed robots to be indistinguishable from humans for plot reasons.
2. We do 99% of our tool use with our arms and hands. We are already very good at building robot arms. We are getting better at robot hands. We can build robot legs, but they're very expensive and they pose a major safety risk for the robot itself and surrounding humans (because the robot can fall if there is a failure). For most applications, why not just put biomimetic hands and arms on a rolling base?
Of course, all this humanoid robotics research is still useful because if you can build a fully humanoid robot you can trivially build a torso-on-rolling-base robot. I sort of suspect that most of the humanoid robotics companies already know that the vast majority of their sales will be in that category.
You can still have a humanoid robot that looks very different from an actual human (and most robots from Asimov's novels were of that kind, although one of the main characters wasn't - https://en.wikipedia.org/wiki/R._Daneel_Olivaw).
Ok, so maybe a robot with wheels could solve most tasks, but it would still be severely limited: couldn't climb stairs (which would make it unsuitable as a domestic robot in a house or multi-storey flat), couldn't drive a car, truck or any other vehicle designed for humans etc.
It's 100% a HMI and moving costs to the other end of supply chain.
We can have optimized automation in warehouses/logistics, but if you talk to any site manager you learn very quickly that no one wants any downtime or impact to their operation to introduce new machinery or optimize traffic, etc. If it is not built with that from the start it's very hard to introduce it later on unless there is a very clear deployment path and cost structure.
And boy, robotics currently has any of those today. Sure, move those billions in to R&D. Time will tell.
Sure, there are already robot vacuum cleaners, but I somehow fail to imagine how (a) specialized robot(s) that can e.g. sort your used laundry, wash it, dry it, iron it (if necessary), fold it and put it back into the cupboard would look like?
Well we got well past washing it by hand in a river. I don't think folding it and storing it in a robotic friendly container until requested would be such a demanding task with today's tech.
If we're designing magic future devices, why are they separate machines? It wouldn't take a robot, I'd just dump my dirty clothes in one end of the machine, and the other end of the machine is my dresser drawers. We can't even handle smushing the washing machine and dryer into one machine, and that technology exists today.
And the other thing is, why is the washing machine itself not seen as a specialized robot? Like if you're designing a machine to machine machines, doesn't it make sense to revisit the whole thing?
Through the first 10 figures I looked at, there was literally A SINGLE ONE that didn't have obvious functional errors. This was not written by someone who read or understood the material. This is slop
Probably a dumb question, because I know nothing about robotics, but:
> The "Zero RPM" Problem
> When a robot bends its knees to stand, the motor must constantly fight gravity. There is no skeletal structure to lock against. To an electric motor, holding a static load—known as stall torque—is the most punishing state possible.
Why not just add some kind of brake that can fully or partially lock the joint?
Brakes are being added to actuators but they're more useful for static holds / locking than dynamic balancing. Standing even in humans is a dynamic balancing activity.
It is, but you don't have to do it at every joint all the time. If you lock your knees your front and back thigh muscles don't have to work on balancing in that axis at all.
It's just weight. If you remove the human scale constraint, you can use hydraulics and easily lock out all your joints with reasonable complexity and maintenance demands. If you're unfamiliar, just look up the history of electrohydrostatic actuators and see where they're at now.
Asking: As gpugreg remarked[0], this is AI slop to the point that it is impossible to trust anything from this article/blog post. As such, I flagged the submission.
- I wonder, is it possible to give a reason to the flag?
- Is flagging the submission without comments the right way to go?
For me, it is important that slowly but surely it goes through that AI slop is not what is accepted here on HN. Yes to have whatever LLM helping with grammar, spelling, etc. but the content should not be the output of a one shot "write me a blog post about humanoid robot actuators" prompt.
I really want to build one of those, they look great fun. (specifically in context of the article I want to see what happens if you lower the CG. Harder balance problem, but might reduce some of the instantaneous torques)
AI was clearly heavily used in the making of this article, and I almost dismissed it as slop. But after reading it I think there's enough correct information here for it to be useful as a general overview of the problems in the space.
I believe that bad/wrong explanations are actually much worse than no explanations.
Many figures seem to be either missing key information (e.g. Fig. 5: the elliptical deformation is not shown - a human artist would have created a very different figure to explain the concept) or plain wrong (Fig. 6: the threaded rollers have the wrong orientation, Fig. 7: the ball is much too large for the bearing and the whole figure seems nonsensical at first glance).
And if the author did not spot these obvious problems with the figures, they either have no clue, accept sloppy work, or didn't even read the article they generated. That article is not really good advertising for the company's products.
(That the link behind the author's name leads to their Wikipedia article which seems to be a revised copy of the CV on their website is interesting, too.)
Its completely impossible to trust an article when half the diagrams have serious issues and headlines are worst slop indicators. I am sure the author is knowledgeable and spend some time on this. But please please, with butter, please, either spend the last 10% to get rid of these issues or just don't publish it! Keep it in your ai research gallery or if you want to publish that gallery as ai research?
Why these screws are used instead of just threads? Because rolling friction is lower than sliding friction. You can use less or more of them trading friction for shock resistance.
I really like the idea of using two series elastic actuators in parallel on the same joint. This way motors acting in opposite direction can pre-tension the springs making the leg stiffer or softer. And if a lot of strength is needed they can act in the same direction summing their forces.
It should be fairly straightforward to control dynamically so you can use pretty much any motor and gearbox.
That's an idea that predates series elastic actuators. It's been used in tendon robots, where there are two winding drums pulling on two cables, with a spring between the end of each cable and the load. The cables oppose each other. Tighten up both, and the joint becomes stiff. Loosen up both, and the joint becomes flexible. Like muscles.
CWRU robotics liked this idea. Lightweight robot arms and painting robots have used it. Snake robots often work that way. Robot hands are often tendon driven.
Tendon robots are not popular for industrial use. The tendon cables wear out. Robots which do the same thing over and over wear their tendons at the same points. Not good in factory settings.
A useful model of a muscle is a spring and damper in parallel, like an auto suspension. The spring's neutral point, the spring's spring constant, and the damper's damping constant are all controllable. An actuator built that way can provide both force and position control, and can absorb shock loads. This is roughly how biological muscles behave.
The most practical actuator like that is a double-ended air cylinder. Each end has two proportional valves, to let air in from the air supply and to let air out to exhaust.
Such devices can be tuned from stiff to rigid, held at any position, and can absorb shock loads, which just compress the air trapped in the cylinder. Festo, the German automation equipment manufacturer, is a leading promoter of that approach.[1] They've popularized precision pneumatic control with a computer managing the valves. Works great in factories where you can plug into a compressed air line. Festo has lots of videos online, if you like looking at pneumatic actuators.
A series elastic actuator is a different concept. It's a stiff spring on the end of a stiff screw-type linear actuator. Shock loads compress the spring, a sensor at the spring detects this, and the linear actuator is commanded to quickly spin up the motor and unload the spring. Force can be measured from the length of the spring. It's a way to fake a real force actuator with a cheap positional one.
This was a popular idea in research robotics, because it's something that can be put together from off the shelf components. It's a useful research tool. But it's not a very good actuator for large loads. There's a race between the applied load squeezing the spring and the motor spinning up to take off the load before the spring bottoms out. It's only useful for a limited load range. If the spring is too stiff, you still take a high shock load at the gear train. If the spring is too soft, the spring bottoms out before the motor can catch up. You can buy series elastic actuators from academic and hobbyist suppliers [2], but they're rarely seen in factories.
There's been much robot actuator progress in the last decade. Not miracles, just money and good specialized mechanical engineering.
I don't know about your city, but my city, while covered in smooth surfaces, still has a lot of interactions between those surfaces. Most of the roads have a curb between the sidewalk and the road, though the ADA means there's curb cuts and ramps. I don't know that I'd agree that they're "perfectly okay". Traversing over a 20 cm/8 inch obstacle constantly seems to make wheels less than perfect.
TLDR: we don't have the actuators required to make humanoid locomotion work reliably.
Also: something every human actually kind of knows. You need to take impacts on muscles, not on mechanical connections. Even if we had the actuators required, you also need perfect control. The only way actuators can work this well is if they properly predict the impacts so that the power of the motor ("the magnetic field") can absorb nearly all the impact. If you try to take the impacts even on human bones (that are very solid and self-repairing) they will break surprisingly quickly.
My opinion is that the need for high reduction is only because we can't have high voltage on the motors. If we either had very small distances between the magnets and electrical wires (think micrometers), or we have voltages in the 100s to 1000s of volts, we don't have to make this poisoned choice. (in a way, VERY small distances between magnets and wires is how human and animal muscles do it. But they go all the way down to sub-10 nanometers)
Naively it feels like the improvements to resistive losses ought to be so dramatic from this that we must already be at some sort of equilibrium position. Double the voltage and divide the resistive losses by 4 - that's neither a trivial gain nor seemingly difficult to achieve? We're not talking kVs here.
The power of money.
I spent some time on legged locomotion back in the 1990s. It was clear then that you wanted torque control, and I did some work on the theory for that, trying to solve it from first principles, not machine learning. Got some nice theory and a patent out. But the parts just weren't there to build such things. As the article points out, the key to this is motor back-drivability. The final drive has to survive shock loads, and it has to dump forces into the motor, where the magnetic fields can take it. As I've quoted before, "you cannot strip the teeth of a magnetic field", a comment from early General Electric locomotive sales. (Locomotives are Diesel-electric, not Diesel with a clutch and shifting gearbox, because the clutch required is huge. Yes, it's been tried.) That's something few areas of engineering cared about, with the exception of aircraft flight control systems with mechanical backup.
Pneumatic actuators looked promising, but proportional dynamic valves were big, heavy, and about $1000 each. Linear motors (not ball screws) looked like the coming thing back then, as 10:1 power/weight ratio had been achieved. But that technology never got much further, and Aura, the biggest player, collapsed in a financial scandal. Series elastic actuators were (and still are) a race between the spring compressing and the ball screw motor starting up. Hydraulics were too clunky; Boston Dynamics built a 400 pound mule, but the Diesel power pack never worked. Direct drive pancake motors were used by some SCARA industrial robots, but those were too big for leg joints. I thought someone would crack the direct drive problem eventually, but nobody ever did. We're still stuck with some gear reduction.
Some of the exotic ideas for muscles mentioned in this article go back that far. The McKinney muscle is old, and not too useful. There was some interest in electrorheological fluids, fluids whose mechanical properties change when an electric field is applied. That didn't become useful either. Shape-memory alloys were a dead end; liquid cooling can overcome the slowness problem, but not the inefficiency problem. Everybody went back to good old electric motors, although they became 3-phase AC instead of DC. It helped that the drone industry made 3-phase motors and their controllers small, cheap, and powerful.
Academic robotics groups were tiny. MIT and Stanford had less than a dozen people each. Progress required hundreds of millions of dollars for all that custom engineering and R&D. The level of effort just wasn't there. Nor would throwing money at the problem prior to machine learning have led to useful products.
It's impressive what's been accomplished in the last five years. It took a lot of money.
Silly question maybe, but didn’t Boston Dynamics have videos of bipedal robots doing acrobatics / running ~7/8 years ago? Kinda looked like they “solved” locomotion then
Their approach required pre-computation and simulation before execution. If you watch their videos carefully, you can see the advance planning work on some of the screens.
Some of it was pre-computed. The middle layer, if you like. The Boston Dynamics group had walking gait of a sort nailed in the 80's; the trotting-on-the-spot that BigDog did was essentially a continuation of those mechanics and that's all based on a conceptually simple balance problem which is intrinsically reactive and not pre-planned. So that's what was going on at the lowest level.
At the top level you have the actual environment, with those meme videos of the robot trotting through a car park, getting kicked off balance, and recovering. The whole point of those tests was to demonstrate how robust their tech was to non-precomputed disturbances.
And between the two you've got the direction and planning layer, telling the robot to go from A to B with some set of suitably convoluted parameters that nobody but the operators would have understood. That planning layer might do all sorts of pre-computation and simulation but it needs to do it in the context of a noisy and possibly adversarial environment. That's equally true for Atlas as much as it was for BigDog, even when there's nobody actually kicking it. What I suspect the precompute and simulation is doing at that layer is a) checking for physical viability of the requested route, and b) parameter tuning in response to sensor readings over a number of runs. Not telling the robot the exact sequence of motions. But I'm nowhere near those teams (oh, I wish) to comment on whether that's true - maybe someone else round here is.
Right. Early thinking on this was zero moment point (ZMP) control. The zero moment point is where, if you land from a stride there, you continue going in the same direction at the same speed. You can speed up by displacing the landing point backwards a bit, and slow down by displacing it further, and turn by displacing it left or right.
When a foot or feet are in contact with the ground, the goal is to stabilize the posture. When in flight, the goal is to hit the zero moment point. Reactive controllers do that. The planner picks the next landing point.
ZMP robots tend to start moving by running in place, then slowly transition to fast forward motion. You see a lot of robot videos like that.
The next step up is to control takeoff. Bend knees, fall forward, launch. Launch angle is a planning problem. Speed and direction changes become much faster. Basic gymnastics are almost within reach. The planner controls launch angle and landing point, and the lower level reactive controls do the millisecond-level control.
It's possible to do all that from first principles of classical dynamics. That level of performance was reached before machine learning. That's what you're looking at in older BD videos.
I can understand pre-computation making the “software” problem of locomotion easier, but how does it help with the hardware problems laid out in the article, ie repeated very high load over a very short amount of time?
BD used hydraulics for a long time. Works, but inefficient. You have to carry the actuators, the tank, probably a hydraulic accumulator, the pump, valves, and the power source for the pump. That's why BD's machines were so big. Someone at Google said "We need to have a conversation about hydraulics", and the dog robot in 2019 was the first all-electric machine.
plenty BD clips of old atlas include oil lines bursting and showing the room with oil.
it's indeed a mess.
BD was under business pressures, and a computerized automaton doing baked ninja back-flips with servos is more impressive than inexpensive FK/IK demos dead-lifting 1000lbs. Google broke that company with their opinions.
Even if private labs have a viable platform solution, people won't care unless they can clone it for free. Not a lot of incentive for design change, but building Kryten 2X4B-523P would be hilarious. =3
overclocking a CPU might make it seem that you solved something and gotten better performance, but sooner or later it breaks down, as I read the article I believe that pre-computation essentially allows you to "overclock" the hardware, and make it seem that you have solved the problem of locomotion when what you have actually done is made something that looks impressive for a very much shorter of time than is usually used to calculate what the hardware can bear.
on edit: apologies if my analogy is not the best.
I feel like the loads would suit electrostatic motors quite well if those could be made appropriately compact.
How can we trust this article or the company if the writer/so-called chief engineer decides to hide himself behind an AI avatar?
From what I can understand this is the Robbie Dickson in question: https://www.huffpost.com/entry/lessons-from-a-serial-ent_b_9...
Nobody has a problem with companies using AI to edit articles, create images. But when even the writer is an AI persona, the trust factor gets destroyed.
You evaluate the merits of the content? From a high-level systems pov, the article is largely correct, even if some details might be missed / simplified
How do you evaluate the merits of a graph of information? I usually read articles like this to learn something, not to grade someone else’s assignment.
Nobody is interested in correct information anymore. In 2026, trustworthy information is what people want.
Yes. This distinction is important.
As a lay person, I have limited knowledge of this field, yet am extremely interested. Unfortunately AI gen content is being used widely to spread spurious information/fake news etc. So my knee jerk reaction to AI gen content is - "this is going to be fake".
If the information you are trying to convey is true, and is technical/objective in nature, then why shy away from associating your identify to it?
I think at very least, the diagrams are AI generated too.
> https://cdn.shopify.com/s/files/1/0554/0567/4694/files/strai... The middle component has teeth on the inside half way round, should be smooth on the inside.
> https://cdn.shopify.com/s/files/1/0554/0567/4694/files/inver... 4 of the 5 orbiting threaded rollers are perpendicular to the screw thread, so wont do anything.
> https://cdn.shopify.com/s/files/1/0554/0567/4694/files/ball-... Ball doesnt fit in the screw thread, just 'squashed' to make it fit?. Screw thread isnt consistent.
> https://cdn.shopify.com/s/files/1/0554/0567/4694/files/stiff... Classic 3 interlocked gears.
> https://cdn.shopify.com/s/files/1/0554/0567/4694/files/optic... Has the elastomer been ripped when the key was inserted?
> https://cdn.shopify.com/s/files/1/0554/0567/4694/files/biolo... Another gearbox that doesnt do anything...
The rest of the website seems filled with just as much slop too...
Right. I'd noticed that the first image, of the harmonic drive, was wrong. The flexible ring gear can't have a disc closely attached; that would kill the flexibility that makes the whole thing go. In some designs, the flexible ring gear is just a ring. The inside of the ring has to be smooth. The elliptical plug is way too small; it has to fit the ring gear.
Harmonic drives are confusing enough. It's amazing that they back-drive at all.
Anything wrong in the text? I'm out of date in this area, and I was reading this as a nice catch-up. What else did they get wrong?
[1] https://www.youtube.com/watch?v=479Xay-Ulrs
I wouldn't be surprised if article is actually AI generated summary from several sources.Guided by human.
It's a new phenomena. Recently got a book on agentic AI which looks like OpenAI docs with added generated water.
Why the AI "engineering expert"? Seems to take some credibility away from what otherwise could be an interesting and informative read.
Yeah I couldn't get past so many issues in the AI generated illustrations. Not useful at all when they are completely wrong.
Even the video is generated. The whole thing is just slop upon slop, I'm amazed that it got to the top of the front page here. I suppose it's a genuinely impressive amount of fakery all integrated together.
right
Check out opensource actuator for robots.
Opentorque actuator
https://www.gabrael.io/new-page
https://github.com/G-Levine/OpenTorque-Actuator
In the context of the torques in the article, a nylon sun gear in the gearbox feels optimistic.
This is AI slop and the article contains some of the worst illustrations I have ever seen. Most do not make any sense mechanically. Here are the worst ones:
- The "orbiting threaded rollers" in figure 6 are not meshing with anything (not that they could, since they are orientated in the wrong direction).
- The ball of the ball screw in figure 7 deforms the screw and the roller screw "meshes" with a flat surface.
- The guy on the pogo stick in figure 14 is jumping himself rather than putting his feet on the stands of the pogo stick.
- In figure 16, a key penetrates the elastomer skin of the optical tactile sensor, destroying it.
- The gears in figure 20 touch perpendicularly.
> - The ball of the ball screw in figure 7 deforms the screw and the roller screw "meshes" with a flat surface.
Really, it is hilarious. This will teach me a lesson or two.
Is it? Now I feel bad for having posted it. I mean, I know some stuff but not enough to judge over all the content in this article. It's unfortunate but from glancing over I thought it was a comprehensive and useful resource. I guess I will just give up on the Internet or something like that.
« Rotary Actuators (The "Reflected Inertia" Trap) », « Quasi-Direct Drive (QDD) — The "Cheetah" Approach «
The pattern ‘something — The « metaphor » <qualifier> ‘ screams Gemini. Gemini seem completely unable to generate a section title that doesn’t follow this annoying pattern.
You just need some AI literacy, this document is very plainly AI generated. You don’t need to be an expert in the content to see that. First time you see an em dash your AI sense should be tingling. Then you see random bolded words alarm bells should be going off. Next you notice 70% of the headlines follow the pattern “The X: Y” and that seals it.
- Figure 3 has "elboly actuators" for the elbow joints (zero hits on Google for the term).
Could it be just the illustrations? I'm not knowledgeable enough to judge the text contents.
Now that you mention it ... apart from the typo the style of the illustration has all marks of AI.
The YouTube channel linked is more of the same. Just the absolute worst slop
Actually, most videos seem to be real, run-of-the-mill showcases of various actuators the company sells. However, the owner seems to have added quite stupid AI-generated preview images lately.
The first video I clicked was an ai generated human reading an ai generated script describing why pulleys are important.
I simply do not understand why you would ever prefer a fully humanoid robot as opposed to a humanoid torso on some other locomotion system.
There are a few of these being sold as products: AGIBOT has some models like that (eg https://www.agibot.com/products/A2_W). One argument that could be made for legged robots is that these wheeled ones can only work in wheelchair-accessible spaces. Legged robots can also balance themselves dynamically: a wheeled robot may tip over if anything violates its static balance, eg. carrying a load high up and going through a steep incline, though I guess having the torso be tiltable as in https://www.agibot.com/products/G2 addresses that.
Legged robots overall have more implementation complexity, spend energy just to idle standing up, but can go over much more varied terrain provided the controller is good enough. There are ways to adapt wheeled bases to different terrains (eg. larger wheels, whegs, RHex, rocker-bogies) but we know how to use legs to locomote over many terrains from personal experience, while the perfect wheeled/non-legged locomotion system perhaps remains to be designed.
There's also the way robotics is going toward data-driven methods, which in some forms (ie. imitation learning) require human teleoperation data. Here having the robot mimic the human form makes the mapping from human joints to robot joints easier (compared to other morphologies where you'd need to figure out how to best approximate a human motion with the joints/joint limits your robot has, though this is not impossible).
Presumably for outdoors or home deployment. The world is designed for bipedal locomotion, and human bipedal locomotion is designed for the world.
But yes, for a factory or commercial environment it doesn't make too much sense. It would be cheaper to adapt the environment, and many commercial environments are already designed to be accessible for wheelchair users anyway.
Isaac Asimov already explained that better than I could (https://www.reddit.com/r/asimov/comments/pm84ud/why_robots_a...):
> Besides that, our entire technology is based on the human form. An automobile, for instance, has its controls so made as to be grasped and manipulated most easily by human hands and feet of a certain size and shape, attached to the body by limbs of a certain length and joints of a certain type. Even such simple objects as chairs and tables or knives and forks are designed to meet the requirements of human measurements and manner of working. It is easier to have robots imitate the human shape than to redesign radically the very philosophy of our tools.
It seems that he wrote that in a book published in 1953, but it's weird, I find, that he was imagining a robot driving a car. I would have thought he would have imagined that cars would become robots well before there would be humanoid robots wanting to drive them. So by the time you have a humanoid robot wanting to drive a car it's just one robot talking to another robot, electronically. And knives and forks are for eating, which humanoid robots presumably don't need to do, and is it likely that humanoid robots will need chairs in the same way that humans do? Altogether, a bad set of examples, I find. Perhaps the thesis would be more convincing with some better examples.
I need to get on to one of the prediction markets.
I’m willing to bet we’ll have a humanoid robot that can drive a car before we have level 5 autonomous vehicles.
And by can drive a car I mean a general purpose humanoid robot that can do basic household chores like move the car and wash it with a foaming brush and a hose.
I don’t mean the robot will be capable of sitting in any regular car and doing level 5 autonomous driving.
So you’re wanting to bet that we’ll get humanoid robots capable of driving a dumb car at L4 before we get cars capable of L5? When we have no humanoid robots driving cars, and many L4 cars driving around? I’ll take that!
Geohot's Comma.ai body is going in that direction, so the question is how much do you want to bet?
Perhaps the samples were chosen specifically as things the audience would have universal familiarity/understanding of thus making his point resonate.
He didn’t craft it for literal interpretation on HN 70 years in the future.
But he was wrong already at the time. The world is made with humans in mind now. As soon as robots that are not human-like become useful enough, the world changes and adapts to them.
Just like the world had no asphalt and curbs before the car. Or just how heavily automated manufacture lines in factories have nothing that looks like a human can interact with it except a few buttoms here and there.
Right, but:
1. Asimov wrote that because he needed robots to be indistinguishable from humans for plot reasons.
2. We do 99% of our tool use with our arms and hands. We are already very good at building robot arms. We are getting better at robot hands. We can build robot legs, but they're very expensive and they pose a major safety risk for the robot itself and surrounding humans (because the robot can fall if there is a failure). For most applications, why not just put biomimetic hands and arms on a rolling base?
Of course, all this humanoid robotics research is still useful because if you can build a fully humanoid robot you can trivially build a torso-on-rolling-base robot. I sort of suspect that most of the humanoid robotics companies already know that the vast majority of their sales will be in that category.
You can still have a humanoid robot that looks very different from an actual human (and most robots from Asimov's novels were of that kind, although one of the main characters wasn't - https://en.wikipedia.org/wiki/R._Daneel_Olivaw).
Ok, so maybe a robot with wheels could solve most tasks, but it would still be severely limited: couldn't climb stairs (which would make it unsuitable as a domestic robot in a house or multi-storey flat), couldn't drive a car, truck or any other vehicle designed for humans etc.
Remember the Segway?
Its predecessor was a stair-climbing wheelchair: https://en.wikipedia.org/wiki/IBOT
One of the codenames for Segway was "Ginger", a reference to Ginger Rogers, because the codename for iBOT was "Fred Upstairs" (a pun on Fred Astaire).
It's 100% a HMI and moving costs to the other end of supply chain.
We can have optimized automation in warehouses/logistics, but if you talk to any site manager you learn very quickly that no one wants any downtime or impact to their operation to introduce new machinery or optimize traffic, etc. If it is not built with that from the start it's very hard to introduce it later on unless there is a very clear deployment path and cost structure.
And boy, robotics currently has any of those today. Sure, move those billions in to R&D. Time will tell.
Why not make a robotic chair? Why not build our environment out of specialized robots instead of using a hammer for everything?
Sure, there are already robot vacuum cleaners, but I somehow fail to imagine how (a) specialized robot(s) that can e.g. sort your used laundry, wash it, dry it, iron it (if necessary), fold it and put it back into the cupboard would look like?
Well we got well past washing it by hand in a river. I don't think folding it and storing it in a robotic friendly container until requested would be such a demanding task with today's tech.
If we're designing magic future devices, why are they separate machines? It wouldn't take a robot, I'd just dump my dirty clothes in one end of the machine, and the other end of the machine is my dresser drawers. We can't even handle smushing the washing machine and dryer into one machine, and that technology exists today.
And the other thing is, why is the washing machine itself not seen as a specialized robot? Like if you're designing a machine to machine machines, doesn't it make sense to revisit the whole thing?
Through the first 10 figures I looked at, there was literally A SINGLE ONE that didn't have obvious functional errors. This was not written by someone who read or understood the material. This is slop
What are the SOTA algorithms for making them walk/move?
I see a lot of videos lately, mostly from China, and I'm curious what everybody is using.
Probably a dumb question, because I know nothing about robotics, but:
> The "Zero RPM" Problem
> When a robot bends its knees to stand, the motor must constantly fight gravity. There is no skeletal structure to lock against. To an electric motor, holding a static load—known as stall torque—is the most punishing state possible.
Why not just add some kind of brake that can fully or partially lock the joint?
Brakes are being added to actuators but they're more useful for static holds / locking than dynamic balancing. Standing even in humans is a dynamic balancing activity.
It is, but you don't have to do it at every joint all the time. If you lock your knees your front and back thigh muscles don't have to work on balancing in that axis at all.
Complexity, weight, failure modes, wear, maintenance, support burden of legacy parts.
Also the brake components are never in the same plane as the drive components, so now you have additional forces to engineer for.
It's just weight. If you remove the human scale constraint, you can use hydraulics and easily lock out all your joints with reasonable complexity and maintenance demands. If you're unfamiliar, just look up the history of electrohydrostatic actuators and see where they're at now.
https://www.moog.com/products/actuators-servoactuators/actua...
You need to keep balancing fast enough. I don't think a break would give you this agility.
Asking: As gpugreg remarked[0], this is AI slop to the point that it is impossible to trust anything from this article/blog post. As such, I flagged the submission.
- I wonder, is it possible to give a reason to the flag?
- Is flagging the submission without comments the right way to go?
For me, it is important that slowly but surely it goes through that AI slop is not what is accepted here on HN. Yes to have whatever LLM helping with grammar, spelling, etc. but the content should not be the output of a one shot "write me a blog post about humanoid robot actuators" prompt.
[0]: https://news.ycombinator.com/item?id=48005917
I also flagged it when I realized the video claimed to be someone it wasn't. Unacceptable in my mind.
Those running blades used e.g. in the paralympics can make locomotion more efficient.
Similar principle taken to more of an extreme: https://is.mpg.de/publications/bb01
I really want to build one of those, they look great fun. (specifically in context of the article I want to see what happens if you lower the CG. Harder balance problem, but might reduce some of the instantaneous torques)
This is a direction worth focusing on. Inspiring work.
I cannot un-see these left border hints, it's driving me crazy.
Does anyone know of a non-slop summary of the same material? Despite the AI diagrams and symbols, the material is more than directionally accurate.
AI was clearly heavily used in the making of this article, and I almost dismissed it as slop. But after reading it I think there's enough correct information here for it to be useful as a general overview of the problems in the space.
I believe that bad/wrong explanations are actually much worse than no explanations.
Many figures seem to be either missing key information (e.g. Fig. 5: the elliptical deformation is not shown - a human artist would have created a very different figure to explain the concept) or plain wrong (Fig. 6: the threaded rollers have the wrong orientation, Fig. 7: the ball is much too large for the bearing and the whole figure seems nonsensical at first glance).
And if the author did not spot these obvious problems with the figures, they either have no clue, accept sloppy work, or didn't even read the article they generated. That article is not really good advertising for the company's products.
(That the link behind the author's name leads to their Wikipedia article which seems to be a revised copy of the CV on their website is interesting, too.)
They should've left out figure 6. it adds very little and the screws are wrong.
Its completely impossible to trust an article when half the diagrams have serious issues and headlines are worst slop indicators. I am sure the author is knowledgeable and spend some time on this. But please please, with butter, please, either spend the last 10% to get rid of these issues or just don't publish it! Keep it in your ai research gallery or if you want to publish that gallery as ai research?
AI that drew the diagram couldn't get it's neurons around how planetary screws in linear actuator should work.
Here's an actual schematic: https://ae-pic-a1.aliexpress-media.com/kf/Sd3fe9841e4ed4871b...
Why these screws are used instead of just threads? Because rolling friction is lower than sliding friction. You can use less or more of them trading friction for shock resistance.
I really like the idea of using two series elastic actuators in parallel on the same joint. This way motors acting in opposite direction can pre-tension the springs making the leg stiffer or softer. And if a lot of strength is needed they can act in the same direction summing their forces.
It should be fairly straightforward to control dynamically so you can use pretty much any motor and gearbox.
That's an idea that predates series elastic actuators. It's been used in tendon robots, where there are two winding drums pulling on two cables, with a spring between the end of each cable and the load. The cables oppose each other. Tighten up both, and the joint becomes stiff. Loosen up both, and the joint becomes flexible. Like muscles. CWRU robotics liked this idea. Lightweight robot arms and painting robots have used it. Snake robots often work that way. Robot hands are often tendon driven.
Tendon robots are not popular for industrial use. The tendon cables wear out. Robots which do the same thing over and over wear their tendons at the same points. Not good in factory settings.
A useful model of a muscle is a spring and damper in parallel, like an auto suspension. The spring's neutral point, the spring's spring constant, and the damper's damping constant are all controllable. An actuator built that way can provide both force and position control, and can absorb shock loads. This is roughly how biological muscles behave.
The most practical actuator like that is a double-ended air cylinder. Each end has two proportional valves, to let air in from the air supply and to let air out to exhaust. Such devices can be tuned from stiff to rigid, held at any position, and can absorb shock loads, which just compress the air trapped in the cylinder. Festo, the German automation equipment manufacturer, is a leading promoter of that approach.[1] They've popularized precision pneumatic control with a computer managing the valves. Works great in factories where you can plug into a compressed air line. Festo has lots of videos online, if you like looking at pneumatic actuators.
A series elastic actuator is a different concept. It's a stiff spring on the end of a stiff screw-type linear actuator. Shock loads compress the spring, a sensor at the spring detects this, and the linear actuator is commanded to quickly spin up the motor and unload the spring. Force can be measured from the length of the spring. It's a way to fake a real force actuator with a cheap positional one.
This was a popular idea in research robotics, because it's something that can be put together from off the shelf components. It's a useful research tool. But it's not a very good actuator for large loads. There's a race between the applied load squeezing the spring and the motor spinning up to take off the load before the spring bottoms out. It's only useful for a limited load range. If the spring is too stiff, you still take a high shock load at the gear train. If the spring is too soft, the spring bottoms out before the motor can catch up. You can buy series elastic actuators from academic and hobbyist suppliers [2], but they're rarely seen in factories.
There's been much robot actuator progress in the last decade. Not miracles, just money and good specialized mechanical engineering.
[1] https://www.youtube.com/watch?v=GedJiaz_E1I
[2] https://www.hebirobotics.com/actuators
Do they run Doom?
Except we don't need 100% bipedal robots. Wheels are perfectly ok for majority of city work and factory floor.
Put the robot on rollerskates break the wheels for the occasional stair.
I don't know about your city, but my city, while covered in smooth surfaces, still has a lot of interactions between those surfaces. Most of the roads have a curb between the sidewalk and the road, though the ADA means there's curb cuts and ramps. I don't know that I'd agree that they're "perfectly okay". Traversing over a 20 cm/8 inch obstacle constantly seems to make wheels less than perfect.
TLDR: we don't have the actuators required to make humanoid locomotion work reliably.
Also: something every human actually kind of knows. You need to take impacts on muscles, not on mechanical connections. Even if we had the actuators required, you also need perfect control. The only way actuators can work this well is if they properly predict the impacts so that the power of the motor ("the magnetic field") can absorb nearly all the impact. If you try to take the impacts even on human bones (that are very solid and self-repairing) they will break surprisingly quickly.
My opinion is that the need for high reduction is only because we can't have high voltage on the motors. If we either had very small distances between the magnets and electrical wires (think micrometers), or we have voltages in the 100s to 1000s of volts, we don't have to make this poisoned choice. (in a way, VERY small distances between magnets and wires is how human and animal muscles do it. But they go all the way down to sub-10 nanometers)
Naively it feels like the improvements to resistive losses ought to be so dramatic from this that we must already be at some sort of equilibrium position. Double the voltage and divide the resistive losses by 4 - that's neither a trivial gain nor seemingly difficult to achieve? We're not talking kVs here.