I've watched and re-watched Aaed's videos on the capstan drive, it's great stuff. High speed, high torque, compliance, effectively no backlash. It's fascinating to watch a legit engineering mind at work.
So, Aaed was our intern last summer - and he’s the real deal. I got to work with him on some really cool bespoke robotic end-effectors and the guy has great design instincts. Picked up mechatronics skills like a sponge, grokked hamming codes in like 30sec from a whiteboard doodle. Super hard worker too, he would stay late in the electronics lab to work on the motor design for what turned out to be CARA - we had fun testing the backlash one evening and ever since I’ve been trying to shoehorn one of those actuators into a project.
Potential and active founders here should consider reaching out (i think a startup setting would suit him better than corporate research), though he’s obviously got his own stuff going on and a degree to finish!
I recently found his videos also. It's one of those things that gets my mind bubbling with ideas for things I want to make, never enough time to do them all though (and this breadboard beside me is asking for attention)
It does make me wonder about the algorithm, Quite a lot of things I find on Youtube turn up on HN a week or two later. I'm not sure if this is an indicator of the effectiveness or failure of the algorithm. It is definitely succeeding in finding videos popular with some people and showing it to more who might share that interest. The question is, are the things I (and consequently many others of similar interests) see the best of all there is, or a subset of the excellent videos out there that happen to get noticed.
I sometimes find channels that are years old with a goldmine of good information. That suggests that there is more good stuff out there than what I see. Were they just unlucky that I didn't see them before? Am I lucky to be seeing them now? It also might be that it is not luck but the algorithm has arbitrarily decided that the video has some special factor that requires promotion or that I have passed some arbitrary threshold of perceived character development that makes me supposedly now interested in such things.
I get absurdly niche videos with 33 views in my feed about robotics but most of them have views in the low thousands. The algorithm isn't holding you back.
The real question: Are the recent recommendations better than previously? Example: If the recommendation engine (presumably driven by some kind of AI/LLM at this point) better understands the content of videos, it can make better recommendations. If anything, it seems logical that smaller channels would begin to appear on your YouTube feed, as long as the content is highly quality and matches your interests.
It's both good and bad: Sometime it suggest some small channels about some niche topics that i found very interesting.
But i don't like how aggressive it is with suggestion. If i search some random topic, my newsfeed will be flooded with all videos about that topic many days later. So sometime i have to use different profile just to not "contaminate" the algorithm.
IIRC he uses a pretty simple algorithm. I remember him discussing the gating mechanism and how he had it follow a cycloid. I think there's a lot of opportunities for others to optimize the algorithms and he is focused on the physical engineering side. I'd love to see him collaborating with someone who does more reinforcement leaning. I also think it's very impressive what he achieves with such simplified algorithms.
If I'm misremembering or missed something please correct me. I'm out now but I'll try to find the video of him talking about gating when I'm back if someone hasn't already linked it.
Also, I love how YouTube has all these "small" creators doing extremely impressive stuff. It's a real shame the algorithms make discovery challenging. The beauty of something like YouTube is not about just getting something to watch, it's by being able to get access to any content. Search is always a difficult problem to solve but I'm afraid it's currently over optimizing for views rather than intent. Which, to be fair, is much harder to measure. But I say over optimized because frequently I can search the title with 90% accuracy and fail to find the video. Something minor like missing an "s" or something effectively non meaningful. It's extremely frustrating...
the algorithm in question is youtube's for how its getting to all the HN users at about the same time. i was also suggested this same video a few days back
My post discussed two algorithms: the gating algo for CARA[0] and the search algorithm for YouTube.
If I'm reading your comment accurately, you're mentioning the discovery algorithm, which is neither of these. I also got the video, but I was already subscribed (it was suggested to me when it was released). Yes, the discovery algorithm has some of these issues but I'm more understanding of that because it's a much harder problem.
Both have self inflicted problems and I think they can be more easily addressed:
Discovery over optimizes to recent views and can get stuck in certain genres[1]. There is also a strong preference to things average user enjoys which doesn't work well for those of us who are only slightly less schizophrenic than the algorithm itself. Too much exploitation, not enough exploration (I wish this was a setting I could adjust. My mood changes, how can I let the algorithm know?)
Search has two critical self inflicted problems.
1) after about 5 results it will suggest completely unrelated videos (looks like it hands off to the discovery feed). Sometimes I need page two... just fucking show me more...
2) the problem I mentioned previously, where it distrusts you prioritizing popular videos over a trivial spelling or grammatical error. Google search has this exact same problem.
[0] my dumbass didn't check which video was linked. It's this one where he discusses it. At 12:30 in the video
[1] this leads me to having tons of YouTube tabs open as I'm unsure if a video I'm interested in but don't have the current bandwidth for is never going to be shown to me again
The comment you responded to clarified that there are two YouTube algorithms (in reality there are many, but was distinguishing the categories). I was talking about search while the response was talking about discovery (like what you see when you open YouTube. The automatic suggestions)
He's easily one of my favorite content creators. Ofc, there are much better engineers, domain experts or more entertaining people on youtube, but he strikes a very enjoyable balance.
I wanted to start writing a list of other tech related, pop-sci and industrial-design Youtubers I kinda enjoy, but noticed just how many channels I'm subscribed to... If there's any interest, I'll drop it, just tell me. Meanwhile I have some filtering and sorting to do.
Hey, I'll reply later to noman-land's comment below, but progress is kinda slow. It takes more time than expected to categorize this many channels and I stopped yesterday after having sorted a bit over 500 and have barely reached the letter "H" (most not tech related ofc, so no worries about getting a monstrous list).
I got a bit more time today, so maybe today evening (German time), elsewise I'll have to skip Friday, so likely Saturday afternoon.
"back in the day", we used capstans to drive film (movie) rolls around the scanning aparatus. Both high speed and precise without backlash. Great stuff. Somehow I always thought maybe lack of high torque is the issue more people aren't using them or wear and tear.. but, apparently not?
Dynema makes a huge difference, a true game changer in many ways. I came across it rigging sailboats where is started out replacing ropes and now replaces steel cables as well. Spectra is another brand with different, equally amazing properties.
Spectra and Dyneema are both UHMWPE. They're different brand names (from Honeywell and DSM respectively) but chemically and mechanically the same, possibly with some small and insignificant differences due to manufacturing processes.
That’s an answer straight from an LLM methinks and not reflective of actual experience. Read up on real world tests for the practical differences. They are very similar but far from “the same”.
I haven't watched the one about the dog, but the one with the initial explication of capstan drives [0] was excellent. I've been dreaming about it for the last year, especially since about the same time another person started working with the da Vinci robot actuators which use cables to generate find motion.
Improved materials. Kevlar and Dyneema.[1] Dyneema is about 15x stronger than steel per unit weight. Kevlar toothed belts have been available for many years, and can be used for many of the same applications as this capstan setup. Neither material has much elasticity.
The advantage over gears is that overloads are distributed over much more material. You don't snap gear teeth. This is good for leg landing shocks.
The relevant feature of Dyneema is specifically that its Young's modulus is much higher than most other organic fibers, about half that of steel, which is probably what you mean by "neither material has much elasticity".
Steel cables would work just as well if weight isn't a consideration, but I think Dyneema is likely to be more resistant to abrasion. However, heat produced by any significant dynamic friction will ruin it immediately, as I found to my sorrow. Kevlar is much more heat-resistant, and of course steel is more heat-resistant than Kevlar.
He actually used steel cables for his initial prototypes, but steel cables have a problem with rolling fatigue that actually gets worse the larger the diameter...so if a tiny steel cable is not strong enough, you might actually have to change the gear dimensions (and therefore ratios) in order to not fatigue them.
Right. Steel cables have to be oiled internally if flexed frequently. This was a major cause of cable failure in gym exercise machines for decades. Newer exercise machines tend to use synthetic flat straps over pulleys.
I'm quite fond of bowstring as a material for this sort of thing. It usually has other fibres mixed in with it so it's a little more bouncy than raw dyneema but that's minimised if you get thread that's intended for crossbows. It's usually waxed so there's some friction to it when you're manipulating it, and it's also easy to source.
It's very non-stretchy rope material. I imagine with enough force you can bend it away from the set position a little bit but it comes back when you remove the force. Nothing like traditional backlash with gears where you can move it with very little force between two adjacent gear teeth.
It's amazing what he's done in terms of the robotics, and the presentation of it to the viewer. I'm amazed at the quality of cinematography on the internet these days.
The implications of the tools we now make available for use in our own personal workshops are still being discovered, and will be for some time.
DIY and "the Internet" (the web?) have fairly little overlap, aside from learning resources existing. Those learning resources could just as well exist, and probably better or more, in a healthier web.
I don't think contrasting the 2 things is making much sense.
I watched this last week and my jaw was on the floor. He's both a great technician, and he has the personality to make it interesting. He walked through his testing strategy far enough that you could understand his methodology and the thought process behind it, but didn't belabor it by making us watch it all. Banger!
I'm going to tell my 12 year old that when he leaves education he wants something like this on his personal web site:
"CARA (Capstans Are Really Awesome) is my latest quadrupedal robot, following ZEUS, ARES, and TOPS. Built over the course of a year, CARA is easily my most dynamic and well-designed quadruped yet."
Why would you put someone with clear talent at building stuff in charge of running a startup? He'll get bogged down in lawyers, day to day operations and growth strategies.
Hire him and put him in R&D in some robotics company.
Y’know, sometimes smart, dedicated, curious, self-directed people are that way with a lot of things.
I was a professor for a long time. My observation was that often a top researcher was also a top teacher and even a top administrator. There are exceptions of course. But if someone is smart and effective at using their attention, those skills transfer to many things.
It’s a pain in the ass when allocating university roles. I want that person to do EVERYTHING ‘cos they always deliver.
On top of studying engineering at uni, his "side-gig" is being creative, empathetic, and fantastic at communication - and your prime recommendation is to "hire" him to be a specialist hidden away in the back office? Which interwebz forum are you on?
EDIT: My last question is clearly an echo-chamber statement. But that doesn't subtract from the fact that, yeah, should he found a business, yes, he'll deal with certain "BS". That is the weight we'll all carry. But he's quite likely capable of moving civilization forwards, so... :shrug:
For many, no; taking on lots of sponsorships, you can make a good amount of money (especially the ones where you agree to do X posts across Y social media accounts for Z period of time, essentially being fully sponsored across a large swath of life).
But for a lot of tech/engineering channels, it'd be immensely difficult to make the same salary as you could working at a FAANG or the like. (I'm making about half what I made when I had a W-2, but it's enough).
Exactly; especially in terms of stress—which is a major problem with my chronic illness, and both the main reason I'm happy to be away from a W-2, but also the thing that makes me most nervous as I'm dependent on marketplace healthcare (which is quite expensive for much worse benefits), and it's harder to get things like life/disability insurance as an independent.
Being a top YouTube producer is not without stress : Reacting to the change in number of followers, responding to comment that can be unpolite, managing sponsors...
I think he'll have success with youtube/vlogging more than getting into the corporate world honestly - especially with some healthy sponsorships and great projects like this.
What is the power consumption of these robots? I often wonder how limited and viable autonomous robots really are. When I look at Tesla's Optimus or Boston Dynamics' spectacular robots, how quickly do they need to be recharged?
My understanding is that those motors he uses are pretty special. And I would expect “efficient” to be part of that special. You’re optimizing for torque and accuracy per unit of mass and energy in this sort of space. I know he talked about them in earlier videos but I no longer recall the details.
Aside from the onboard computing required to power the AI model controlling the robot, which also consumes energy, it is important to be as efficient as possible. However, if an Optimus is intended to replace human labour, such as lifting objects and wander around with it, the energy efficiency of the platform becomes negligible.
The platform in the video and the robot dogs from Boston Dynamics are ideal for tasks where they are only limited by their own weight and the amount of computing power required to navigate, such as exploration.
I suppose that's why we mostly see autonomous delivery robots on wheels.
Or maybe I am being too pessimistic about other platforms...
Breaking Taps on YouTube did a really awesome video on a somewhat similar mechanism (I'm no mechanical engineer haha, it was new to me!), rolling contact joints. I love the idea of using string/ropes. Worth checking out as well if this kind of stuff interests you! https://www.youtube.com/watch?v=TQiLLcumqDw
Well, it probably wasn't that much effort. When you're 3D printing you're going to end up printing everything 2-3 times anyway, so why not dial in the ratio while you're at it?
And you can't really declare your design is "high precision" and present yourself as someone others should take transmission design advice from if you aimed for a gear ratio of 8 and achieved "somewhere around 7.9 to 8.2"
It probably doesn't matter so much whether it's 7.913 or 8.186, but it would be important to know the exact value for kinematics. One way to do that is to build an object very accurately, the other is to build inaccurately and then measure the result after the fact.
It's also interesting because competing actuators with strain-wave, cycloidal, or planetary gearboxes will state exactly what the ratio is. The actual gear teeth may not be spaced out perfectly around the circumference, but the number of teeth is an integer with an infinite number of zeros.
Yeah, I think one of the nice things about making it a "clean" number (either an integer or a rational with a small integer denominator) is that you can easily validate it without needing high-precision measurement equipment: put a mark on both gears (maybe even embedded in the 3D print), line up the marks, rotate the large gear 1 full rotation, and count the number of rotations the smaller gear makes. Check to see if the marks line up perfectly after those rotations.
It could though. I don’t think the thought has occurred to him yet. He could make the stack twice as high and go around twice as far. If you moved the worm gear you could go farther, but I don’t know how he would do that with his drive.
He could also go with narrower rope, and spread the load over more windings, which would give him more throw.
What are the loads on a drive in this use case? 35lbs robot bouncing on one foot would be 350lbs-ish of dynamic load? That can be handled with 2mm Dyneema.
> Also, I wonder how resistant this mechanism is to wear and fatigue.
He actually discussed this in an earlier video for his initial tests on the capstan drive. He ended up testing the rope he used for around 358 hours (two weeks) on continuous use in the drive itself with very low backlash
Because when a real engineer puts 2 and 2 in and gets 3.8 out, it vexes them and they want to at least know why they can’t get 4. He’s trying to make a machine that does what he told it to do, so that he understands what is actually happening.
I think it's about kinematics, the more precise your gears the better the model fits the real world.
That's why pro crews don't use gears and ropes. At high impulses deformations and elasticity throw the kinematics off what's actually happening. Modeling the deformations and the elasticity is a computational no no. Instead what you see is the motors right on the joints.
At least that was the case last time I had a look at robotics.
> That's why pro crews don't use gears and ropes. [...] Instead what you see is the motors right on the joints.
The answer here, as with so many things in robotics, is: It Depends.
UR10e robot arm that can lift a 4kg object with a reach of 1m and has sub-1mm repeatability? Strain wave gears in the base and shoulder joints, 100:1 ratio.
MIT Mini Cheetah robot dog that can do backflips? 6:1 planetary gearbox.
Shadow Hand with 20 degrees of freedom? Tendon driven, with the 20 motors in the forearm to keep the fingers slim.
Little dinky Huggingface SO-101? Servo motors, integrating 1:345 gearing with a series of 6 tiny brass gears.
Mid-price CNC milling machine, if you call that a robot? Really long ballscrews, driven by stepper motors.
Surely you mean driven by closed-loop servos with encoders, don't you? Jacques Mattheij wrote a long post about how he ended up having to replace all the CNC machines he sold that used stepper motors because he didn't know any better at the time, and brushless motors are a lot faster and more powerful as well as not losing steps.
I guess "mid price" isn't the most helpful term when machines vary from $100 to $100,000 huh?
I think the X-Carve 3-axis wood carver uses stepper motors with belts of all things. The Shapeoko Pro is leadscrews and stepper motors. Wazers, I believe, are belt-and-servo driven. And a lot of 'CNC conversion kits' you can order online use stepper motors. Plus of course laser cutters have really low torque requirements, so they've got a lot of design freedom.
Arguably those are "cheap" rather than "mid price" it just felt weird to declare a $4000 machine to be "cheap"
Actually, it was the Shapeoko 1 which introduces the idea of belts (MXL), then switching to Gates GT2 belts with the SO2 (ob. discl., I got a free machine for doing the instructions, see Github). Since it was opensource, the X-Carve was forked from the SO2, though the X-Carve Pro was a completely new design.
Note that belts were continued through the Shapeoko 3 (I got a machine as a "thank you"), though the Z-axis got a leadscrew in the Z-Plus upgrade, then the Pro (since, I got a job with the company and got an XXL as part of my employment), then the 4 (the original Pro is referred to as a 4 Pro sometimes), and it is only with the 5 pro that ball screws were switched to for all axes (and I now have a 5 Pro).
For an example of what a belt-drive CNC can do see:
On the other hand, a 100:1 gearbox gives you much-needed torque if you're lifting a load at the end of a long arm, it makes your encoder 100x more precise (in terms of repeatability) and it makes your motor brake 100x stronger.
Back-drivability is the enemy of precision, so many robotic applications can do without it.
Almost all modern servo driven units I've seen prefer to allow some compliance in the end effector. The UR5 and UR10 series for example can use force limiting control loops, and are safer to use around people.
The old "fast, cheap, or good... choose any two joke is mostly still true. =3
The UR10 uses 100:1 strain wave gearing in the base and shoulder joints.
You’re right that it has a freedrive mode, and force control modes. But it’s a rigid, low-backlash robot with the compliance achieved in software afterwards.
Expensive, naturally, but none of the problems that come with things like series elastic actuators.
> At high impulses deformations and elasticity throw the kinematics
Sure. Yet evolution has achieved astonishing kinematics with all manner of deformation and elasticity inherent to the materials, and also constantly changing physical properties, using low resolution data. We cannot build permanently lash free mechanical devices at reasonable cost and reasonable size/weight. Eventually, the answer must be pervasive real-time compensation throughout the kinematic model.
> Modeling the deformations and the elasticity is a computational no no.
Why? Nervous systems do this. That's why you can change your shoes and still walk upright.
In the video he mentions how the specific type of Dyneema cord he's using is well-suited for the application (compared to other kinds of rope/cord). It's particularly strong, light, and inelastic; a lot of climbing equipment uses a version of it for similar reasons.
> Modeling the deformations and the elasticity is a computational no no. Instead what you see is the motors right on the joints.
That sounds like “It’s not wrong, we just don’t do it”. There are some amazing examples of imprecise drive systems compensated for by excellent control systems all over the world, for millions of years.
Yet for some reason unknown to me, people get annoyed when you tell them „let us meet at 13:37. that is no more accurate than let us meet at 14:00 hours”
I don't think the number of the gear ratio really matters, what matters is that you know what it actually is (since every IK calc depends on said ratio); 8:1 is probably arbitrary and/or looks nice & might simplify some stuff.
It might be a lot easier to check the ratio "by hand" (by counting rotations etc) if it's numerically simple. (IIRC in some earlier videos he noticed that the pulley size ratio wasn't producing the expected movement ratio, because they were built as an obvious 8:1 or 10:1 or something, and didn't match - which led to him figuring out the subtleties of the design - I can easily imagine wanting to preserve that aspect just for debugging, at that point, even if you now have correct math.)
Thanks for sharing this! What a treat of a video. It's a fun project, and it's presented very well. This guy has a talent for communication - the video was super clear and well explained. I really admire that ability and I want to get better at it.
Also the Spanish words for "face" (no me gusta tu cara, boludo) and "expensive and female" (esa computadora es re cara, boludo). https://en.wiktionary.org/wiki/cara tells us it also has meanings in Aragonese, Asturian, Catalan, Crimean Tatar, French, Galician, Indonesian, Italian, Latvian, and another couple dozen languages.
There is a weird mixture of hope and dread in me as I watch this. I am ridiculously excited over a person like me being able to mess around with something that, until recently, was gated behind, well, a lot of hard problems to overcome ( I am only now slowly getting through old Peter Scott's robotics to get some perspective ). By comparison, it should be so much easier to explore aspects of robotics that may go beyond strict math and engineering.
Gated how? It's all on textbooks, industry standards, Internet, everywhere. You could buy a KHR-3 without even driver's license for $1.5k cash for past ~20 years. I guess CUDA was technically gated behind "I swear I'm not a bad guy" button, but that was it.
I'm surprised that Ukraine isn't DIYing full-on solid motor cluster SAMs and armor piercing ATGMs. I thought those kinds of devices were something just about any sufficiently developed country can do in days to weeks should such national emergency arises and all bets were off, except nations in peacetime has moral obligations to do no evil.
Land mines have been killing people by surprise for decades. Flying drones, some of them dropping land mines and others lying in wait for incoming convoys, produced 70% of the casualties on both sides in the Ukrainian invasion last year, and a few weeks ago the Ukrainians shipped some drones to near a Russian airport several thousand kilometers from Ukraine to blow up a bunch of airplanes. This betokens a future of borderless warfare in which probably most of your family members will be killed over the next few decades, by one or another type of semi-autonomous or fully autonomous weapon.
CARA is a super cool project which is never going to kill anybody, but it's another piece of evidence that the cost of the technology for such weapons has decreased enormously.
That said, talking about the dread is going to get boring fast, because nearly every story on the HN front page is catapulting us toward that future.
Good point. If we do press panic button each time, the word will lose its meaning. Naturally, it does not help that this particular community did happen to catapult us towards the future a fair bit over the past years.
I'm more optimistic. I think filling battlefields with autonomous weapons will make humans obsolete there and reduce war causalities. Basically any attack on humans is going to be a war crime and terrorism akin to attacks on civilians russia is doing right now
Well, suppose you have two countries at war, say, Burnia and Hevonia. They each produce a million autonomous weapons.
Burnia sends their million autonomous weapons to kill the million people they judge are most crucial to Hevonia's war effort, prioritizing Hevonia's political leadership and military officers.
Hevonia, meanwhile, sends their million autonomous weapons to destroy Burnia's autonomous weapons, when they can find them, but not to attack any humans.
Who wins the war?
I think Burnia does, because even if Hevonia's weapons are 99% effective, Hevonia's government has still lost its top ten thousand people, including all of their military officers, while Burnia has only lost half a billion dollars. That's going to make it impossible for Hevonia to keep fighting. And I think 50% effective is more likely.
It's the same question as when one opponent is ready to nuke enemy cities while the other is not. In theory everything turns out the same way you postulate. In practice cost of winning the war in this manner might be higher then the cost of losing localized, limited conflict.
Who cares if it's weird? Why are you saying "weird" as if that were a bad thing? It's just a derogatory synonym for "unusual". A take would have to be weird to be a contribution to the conversation, because if it's not weird, it's a widely held opinion we all already know, and nobody gains anything by reading it.
What matters, given that it meets the minimal bar of weirdness to be potentially worthwhile, is whether it's correct. Which depends on what future combatants will do once weaponry actually is autonomous, not what Ukraine and Russia are doing with remotely-piloted FPV drones.
I'd call it delusional than weird. US is clearly controlling Ukrainian churn rate so that Russian military and Ukraine as nation burn down at the same rate into disappearance so that the postwar Ukraine can be rebuilt with more Western leaning meatbags from surrounding nations. And some of people here are framing exactly that as utopian bloodless fights of machines. Calling it weird is itself almost weird.
I think we are in transition period and nations are just learning that sending people to contact line is a terrible idea. People never learned things like that very fast. It often takes a generation.
> This betokens a future of borderless warfare in which probably most of your family members will be killed over the next few decades, by one or another type of semi-autonomous or fully autonomous weapon.
You're talking about a whole bunch of targeted and intentional attacks in a literal warzone, the most lawless kind of place on the planet and you're complaining that they are hitting their intended military targets?
Meanwhile Gaza is boring to you, because Israel uses conventional bombs with humans in the loop, even though the collateral damage matches your prediction today.
I'm not complaining, exactly; I'd rather the Ukrainians be blowing up Russian planes, which were thousands of kilometers from the literal warzone, than the Russians blowing up Ukrainian planes thousands of kilometers away.
I'm saying that the events we're seeing in the Ukrainian war are convincing evidence that the nature of war has changed, and the implications of that change for human society are disquieting.
For two million years, people have often resolved conflicts by warfare. But that warfare, though it has never ended completely, has always been localized, which meant that most places most of the time were peaceful.
This is the foundation of the system of international relations in which different states exercised monopolies of legitimate violence over geographical territories: by so doing, they could prevent warfare and provide security. This system has existed to one extent or another for thousands of years.
Warfare is no longer localized, so now the only place for anywhere to be peaceful will be for everywhere to be peaceful. It is no longer possible for states to provide security to the people within their borders. Consequently statehood itself has lost its meaning. The last time we saw such a change in the nature of warfare was twelve thousand years ago when the first states arose.
I have some corrections for you about other things that you did not understand either about my comment or about the wars we are commenting on.
It bears repeating that the planes blown up in Operation Spiderweb were not in a literal warzone, nor anywhere close to a literal warzone. Some of them were thousands of kilometers from the literal warzone, in a part of Russia that has China and Mongolia between it and Ukraine.
The Ukrainians reportedly did have humans in the loop for that attack in particular, and both the Ukrainians and the Russians are mostly using FPV drones rather than autonomous drones of any sort at this point, unless we count landmines as "autonomous drones". In https://www.youtube.com/watch?v=RmfNUM2CbbM Russian drone developer Sergey Tovkach specifically says that Starlink is the reason Ukrainian sea drones have established control of the Black Sea — Russia could build sea drones but does not have an equivalent to Starlink, so evidently both Ukraine and Russia are dependent on remote piloting for their sea drones. He also says Starlink's latency is too high for piloting quadcopters. (Evidently Operation Spiderweb used the Russian cellphone network.)
The Israeli attacks are also using precision-guided weaponry and (in some sense) AI, not conventional bombs, and they're killing lots of innocent people, but you're right that it's less worrisome to me, for three reasons:
- The total number of people they're killing is about an order of magnitude smaller, about 80,000 so far (out of a total of 2.1 million Gazans), versus roughly a million in the invasion of Ukraine. Many more civilians have been killed in Gaza, but both Russia and Ukraine practice conscription, so it's not as if the soldiers being killed are only or even mostly volunteers; they were civilians until being, in many cases, conscripted.
- The weapons Israel are using are mostly very expensive, which limits the number of people they can afford to kill with them. By contrast, the drones being mass-produced in both Ukraine and Russia cost only a few hundred dollars each, and each country is expected to produce about 3 million of them this year.
- The people Israel are killing are almost entirely in close geographical proximity to Israel. Gaza City is 80 km from Jerusalem. Beirut, the residence of many of the Hizbullah personnel that Israel killed with explosive charges in pagers (along with, in several cases, their children), is 230 km from Jerusalem.
Therefore, while Israel's war is clearly causing terrible suffering to millions of innocent people, and it could easily spiral into a Third World War, it does not represent the advent of a new, borderless mode of warfare in which the cost of untraceably killing a precision-targeted human anywhere in the world is similar to the cost of a small air conditioner or 20 kg of beef.
The developments we are seeing of the mode of war in the Ukrainian invasion do represent such a transition, and that is true regardless of how the conflict goes. But mostly it is not the Ukrainians and Russians who are responsible for the transition; it was an inevitable result of developments in batteries, motors, power electronics, and 3-D printing and (more broadly) digital fabrication.
What will the new equilibrium look like, now that statehood is effectively meaningless? Well, it might be better or it might be worse, but not many people currently alive will live long enough to find out.
Fascinating! I want to get into this type of stuff. But I have no idea where to start, I just have just a CS degree and 3 years experience as a developer.
I recommend a Brachiograph build. It will introduce you to some fundamentals of PWM and inverse kinematics. It is well documented but not cookie-cutter. Using a Raspberry Pi will give you more direct access to running the servos than the microcontroller experience. All the parts are infinitely reusable afterward if you don't want to keep it around.
Sample Supply List for $80 budget:
Pi Zero with header $20: https://www.adafruit.com/product/6008
Power supply $9: https://www.adafruit.com/product/1995
SD Card $10: https://www.adafruit.com/product/1294
Three hobby servos $18: https://www.adafruit.com/product/169
Breadboard wires $5: https://www.adafruit.com/product/153
Breadboard $5: https://www.adafruit.com/product/64
Glue, popsicle sticks, pen and paper $10
Arduinos and hobby servos. No, neither of them are "industrial grade" and yeah, you'll reach their limits pretty quickly, but building a physical thing that does stuff is (in my experience) a huge motivator.
Or if you're already all over the basics, figure out what kind of stuff you want to build and then try and build it. :)
To begin with you can get a cheap robotic kit from Amazon, there are many of them, and put it together. That's probably the easiest and fastest way. From here you can read more about servos and controllers, modify its mechanics and software.
New here and just stumbled upon this. Seen these "robodogs" live in Vegas mining conference. The usage is picking up, their functionality and range of movement is so much more advanced now, and the list of actions they can complete much more complex. Awesome stuff!
Very cool robot dog and interesting video!
Can the dog climb stairs? Isn't capstan drive temperature sensitive, e.g. the ropes will be shorter in cold and longer in warm wheather?
I remember a post previously related to him here on HN but I am surprised at how I forgot about him and how cool he is to build all these incredible stuff and also teaches in his videos. I am subscribing so that I never miss anything from him again.
What a horrible video is this, with the robotic translated AI voiceover?
Update: Ah, weird, if I watch the non-embedded one on youtube it is the original in English with normal sound. It's the one embedded on his web site which has AI translation to German.
That's not the video's fault, but YouTube's. YouTube has started adding automatic AI voiceover translations to all videos, unless the uploader explicitly disables them for each single video. It's opt-out and the option is well-hidden, so most uploaders will not disable it.
As an end user, there's nothing you can do to prevent from seeing them. But you can change the audio track to the original while the video is playing.
How do we make such robots intelligent? Like you only tell it to learn to jump the rope and then it goes through trial and error(s) to figure out a way to do it, and if it can’t and needs a physical upgrade then it tells you that? (Like a certain gear ratio etc)
It's just a gut feeling but I'd trust a feedback based backlash elimination mechanism more than rope based, especially in the long run and/or with a large deployed base.
Like I said gut feeling. I suppose over time there may be some play in the axles, for example. Something feedback could compensate for but a rope could not.
Anyway good engineering requires extensive testing over many hours of movements.
> Anyway good engineering requires extensive testing over many hours of movements.
He actually discussed this in an earlier video for his initial tests on the capstan drive. He ended up testing the rope he used for around 358 hours (two weeks) on continuous use in the drive itself with very low backlash
> Programming takes the cake for what is both my most and least favourite part of any build. Nothing quite makes you pull out your hair and ask yourself, 'What the heck have I gotten into?' Like spending weeks programming a robot that just won't work. Eventually though, you fix that one line of code that makes all the difference. And then it's smooth sailing. Well, kind of.
I feel this deeply, also this whole video is quality content.
this guy's work is pretty amazing
I whole heartedly believe that I'll own a gang of (armed) quadrapets for home/property/personal protection within the next 10 years. I'm equally worried that once these become commodities they might be worse than an electric scooters have become to us bipeds on sidewalks but on balance I can't wait
It might be fun to optimize the shape of the rolling contact surface of the capstan drive away from cycloid to make it even better suited for specific application, like a robot dog leg that smoothly runs for miles.
I don't think it is fair to downvote this comment. It is a genuine concern and should be addressed. Amusingly, given that the thought has entered mass consciousness by means of question on 'whether all this is a simulation', matrix ( the animated series ) explored this question a little and it is interesting in how the timeline aligns with what the movie presented.
It's absolutely a legitimate concern but since this robot project has no chance of replacing humanity, it seems like a derail in terms of the subject matter of the video. It's as if someone posted a link to an incremental advancement in graphene aluminum-ion battery tech and a person chimed in with a general lament about heavy metal battery toxic waste disposal.
That's an interesting reference, but do our timelines really align with what was presented in (ani)matrix? Their timelines are quite murky, but it seems that in their universe things only started getting problematic between humans and AIs around 2090 and The First Machine War started in 2139 [0]. With how fast things are progressing, it's not clear to me that we have so much time.
On the flip side, capitalistic / private / special interests both controlling the progress and having the most ability to utilize it to further centralize power and wealth is deeply concerning. We can already see more controversial figures involved in AI using it to spread their personal viewpoints.
It feels really easy to see how our jobs/labor and therefore our capital and therefore our value in the modern system are being directly attacked by these capabilities and deeply concerning to imagine how further centralization of power could be good for the masses.
When I came across this amazing project and wanted to share it to HN, I was debating whether to post the youtube link or the project page. I decided to post the project page and mention the youtube link in the description for those who prefer video, but somehow that description got posted as a comment instead (not sure how that happened?). Anyway as you said the video is embedded in the project page so it wasn't really necessary
I've watched and re-watched Aaed's videos on the capstan drive, it's great stuff. High speed, high torque, compliance, effectively no backlash. It's fascinating to watch a legit engineering mind at work.
So, Aaed was our intern last summer - and he’s the real deal. I got to work with him on some really cool bespoke robotic end-effectors and the guy has great design instincts. Picked up mechatronics skills like a sponge, grokked hamming codes in like 30sec from a whiteboard doodle. Super hard worker too, he would stay late in the electronics lab to work on the motor design for what turned out to be CARA - we had fun testing the backlash one evening and ever since I’ve been trying to shoehorn one of those actuators into a project.
Potential and active founders here should consider reaching out (i think a startup setting would suit him better than corporate research), though he’s obviously got his own stuff going on and a degree to finish!
I recently found his videos also. It's one of those things that gets my mind bubbling with ideas for things I want to make, never enough time to do them all though (and this breadboard beside me is asking for attention)
It does make me wonder about the algorithm, Quite a lot of things I find on Youtube turn up on HN a week or two later. I'm not sure if this is an indicator of the effectiveness or failure of the algorithm. It is definitely succeeding in finding videos popular with some people and showing it to more who might share that interest. The question is, are the things I (and consequently many others of similar interests) see the best of all there is, or a subset of the excellent videos out there that happen to get noticed.
I sometimes find channels that are years old with a goldmine of good information. That suggests that there is more good stuff out there than what I see. Were they just unlucky that I didn't see them before? Am I lucky to be seeing them now? It also might be that it is not luck but the algorithm has arbitrarily decided that the video has some special factor that requires promotion or that I have passed some arbitrary threshold of perceived character development that makes me supposedly now interested in such things.
I get absurdly niche videos with 33 views in my feed about robotics but most of them have views in the low thousands. The algorithm isn't holding you back.
i noticed recently YT has been suggesting micro-small channel(<1000 subscriber) to me, not sure if this is good thing.
The real question: Are the recent recommendations better than previously? Example: If the recommendation engine (presumably driven by some kind of AI/LLM at this point) better understands the content of videos, it can make better recommendations. If anything, it seems logical that smaller channels would begin to appear on your YouTube feed, as long as the content is highly quality and matches your interests.
It's both good and bad: Sometime it suggest some small channels about some niche topics that i found very interesting.
But i don't like how aggressive it is with suggestion. If i search some random topic, my newsfeed will be flooded with all videos about that topic many days later. So sometime i have to use different profile just to not "contaminate" the algorithm.
If I'm misremembering or missed something please correct me. I'm out now but I'll try to find the video of him talking about gating when I'm back if someone hasn't already linked it.
Also, I love how YouTube has all these "small" creators doing extremely impressive stuff. It's a real shame the algorithms make discovery challenging. The beauty of something like YouTube is not about just getting something to watch, it's by being able to get access to any content. Search is always a difficult problem to solve but I'm afraid it's currently over optimizing for views rather than intent. Which, to be fair, is much harder to measure. But I say over optimized because frequently I can search the title with 90% accuracy and fail to find the video. Something minor like missing an "s" or something effectively non meaningful. It's extremely frustrating...
the algorithm in question is youtube's for how its getting to all the HN users at about the same time. i was also suggested this same video a few days back
My post discussed two algorithms: the gating algo for CARA[0] and the search algorithm for YouTube.
If I'm reading your comment accurately, you're mentioning the discovery algorithm, which is neither of these. I also got the video, but I was already subscribed (it was suggested to me when it was released). Yes, the discovery algorithm has some of these issues but I'm more understanding of that because it's a much harder problem.
Both have self inflicted problems and I think they can be more easily addressed:
Discovery over optimizes to recent views and can get stuck in certain genres[1]. There is also a strong preference to things average user enjoys which doesn't work well for those of us who are only slightly less schizophrenic than the algorithm itself. Too much exploitation, not enough exploration (I wish this was a setting I could adjust. My mood changes, how can I let the algorithm know?)
Search has two critical self inflicted problems.
1) after about 5 results it will suggest completely unrelated videos (looks like it hands off to the discovery feed). Sometimes I need page two... just fucking show me more...
2) the problem I mentioned previously, where it distrusts you prioritizing popular videos over a trivial spelling or grammatical error. Google search has this exact same problem.
[0] my dumbass didn't check which video was linked. It's this one where he discusses it. At 12:30 in the video
[1] this leads me to having tons of YouTube tabs open as I'm unsure if a video I'm interested in but don't have the current bandwidth for is never going to be shown to me again
The quote in your intial reply ("It does make me wonder about the algorithm") was referring to the YouTube algorithm.
The comment you responded to clarified that there are two YouTube algorithms (in reality there are many, but was distinguishing the categories). I was talking about search while the response was talking about discovery (like what you see when you open YouTube. The automatic suggestions)
He's easily one of my favorite content creators. Ofc, there are much better engineers, domain experts or more entertaining people on youtube, but he strikes a very enjoyable balance.
I wanted to start writing a list of other tech related, pop-sci and industrial-design Youtubers I kinda enjoy, but noticed just how many channels I'm subscribed to... If there's any interest, I'll drop it, just tell me. Meanwhile I have some filtering and sorting to do.
I'm interested as well! I never can get enough of these builders and engineers.
Hey, I'll reply later to noman-land's comment below, but progress is kinda slow. It takes more time than expected to categorize this many channels and I stopped yesterday after having sorted a bit over 500 and have barely reached the letter "H" (most not tech related ofc, so no worries about getting a monstrous list).
I got a bit more time today, so maybe today evening (German time), elsewise I'll have to skip Friday, so likely Saturday afternoon.
I am interested. I love these builder channels.
"back in the day", we used capstans to drive film (movie) rolls around the scanning aparatus. Both high speed and precise without backlash. Great stuff. Somehow I always thought maybe lack of high torque is the issue more people aren't using them or wear and tear.. but, apparently not?
Dynema makes a huge difference, a true game changer in many ways. I came across it rigging sailboats where is started out replacing ropes and now replaces steel cables as well. Spectra is another brand with different, equally amazing properties.
Interesting you've mentioned sailboats. I've encountered dyneema for the first time in spearfishing!
Spectra and Dyneema are both UHMWPE. They're different brand names (from Honeywell and DSM respectively) but chemically and mechanically the same, possibly with some small and insignificant differences due to manufacturing processes.
That’s an answer straight from an LLM methinks and not reflective of actual experience. Read up on real world tests for the practical differences. They are very similar but far from “the same”.
These drives sound amazing, so why are they not used everywhere? What are the disadvantages?
I haven't watched the one about the dog, but the one with the initial explication of capstan drives [0] was excellent. I've been dreaming about it for the last year, especially since about the same time another person started working with the da Vinci robot actuators which use cables to generate find motion.
0. High Precision Speed Reducer Using Rope: https://www.youtube.com/watch?v=MwIBTbumd1Q
1. Building a DIY Surgical Robot https://www.youtube.com/watch?v=d_8rHKrwr-Q
The "DIY Surgical Robot" vid is so good. And it's also posted over a year ago with no follow-up. Damn.
How is there no backlash? I can’t imagine a rope drive without backlash.
Improved materials. Kevlar and Dyneema.[1] Dyneema is about 15x stronger than steel per unit weight. Kevlar toothed belts have been available for many years, and can be used for many of the same applications as this capstan setup. Neither material has much elasticity.
The advantage over gears is that overloads are distributed over much more material. You don't snap gear teeth. This is good for leg landing shocks.
[1] https://www.impact-fibers.com/info/unveiling-the-strength-ke...
The relevant feature of Dyneema is specifically that its Young's modulus is much higher than most other organic fibers, about half that of steel, which is probably what you mean by "neither material has much elasticity".
Steel cables would work just as well if weight isn't a consideration, but I think Dyneema is likely to be more resistant to abrasion. However, heat produced by any significant dynamic friction will ruin it immediately, as I found to my sorrow. Kevlar is much more heat-resistant, and of course steel is more heat-resistant than Kevlar.
He actually used steel cables for his initial prototypes, but steel cables have a problem with rolling fatigue that actually gets worse the larger the diameter...so if a tiny steel cable is not strong enough, you might actually have to change the gear dimensions (and therefore ratios) in order to not fatigue them.
Right. Steel cables have to be oiled internally if flexed frequently. This was a major cause of cable failure in gym exercise machines for decades. Newer exercise machines tend to use synthetic flat straps over pulleys.
Dyneema is much easier to work with, too.
I'm quite fond of bowstring as a material for this sort of thing. It usually has other fibres mixed in with it so it's a little more bouncy than raw dyneema but that's minimised if you get thread that's intended for crossbows. It's usually waxed so there's some friction to it when you're manipulating it, and it's also easy to source.
It's very non-stretchy rope material. I imagine with enough force you can bend it away from the set position a little bit but it comes back when you remove the force. Nothing like traditional backlash with gears where you can move it with very little force between two adjacent gear teeth.
It's amazing what he's done in terms of the robotics, and the presentation of it to the viewer. I'm amazed at the quality of cinematography on the internet these days.
The implications of the tools we now make available for use in our own personal workshops are still being discovered, and will be for some time.
Feels like the golden age of DIY engineering
I agree 1000%, which is why I find the whole “the internet sucks now” movement really disheartening. There are huge upsides to a network this large.
DIY and "the Internet" (the web?) have fairly little overlap, aside from learning resources existing. Those learning resources could just as well exist, and probably better or more, in a healthier web. I don't think contrasting the 2 things is making much sense.
I watched this last week and my jaw was on the floor. He's both a great technician, and he has the personality to make it interesting. He walked through his testing strategy far enough that you could understand his methodology and the thought process behind it, but didn't belabor it by making us watch it all. Banger!
Amazing presentation! - somebody hire this kid asap
https://www.aaedmusa.com/
I'm going to tell my 12 year old that when he leaves education he wants something like this on his personal web site:
"CARA (Capstans Are Really Awesome) is my latest quadrupedal robot, following ZEUS, ARES, and TOPS. Built over the course of a year, CARA is easily my most dynamic and well-designed quadruped yet."
That would be a terrible path for someone with this extreme level of demonstrated talent, motivation, and follow-through.
Much better for someone to fund a startup run by him.
Why would you put someone with clear talent at building stuff in charge of running a startup? He'll get bogged down in lawyers, day to day operations and growth strategies.
Hire him and put him in R&D in some robotics company.
Y’know, sometimes smart, dedicated, curious, self-directed people are that way with a lot of things.
I was a professor for a long time. My observation was that often a top researcher was also a top teacher and even a top administrator. There are exceptions of course. But if someone is smart and effective at using their attention, those skills transfer to many things.
It’s a pain in the ass when allocating university roles. I want that person to do EVERYTHING ‘cos they always deliver.
... what tf?
On top of studying engineering at uni, his "side-gig" is being creative, empathetic, and fantastic at communication - and your prime recommendation is to "hire" him to be a specialist hidden away in the back office? Which interwebz forum are you on?
EDIT: My last question is clearly an echo-chamber statement. But that doesn't subtract from the fact that, yeah, should he found a business, yes, he'll deal with certain "BS". That is the weight we'll all carry. But he's quite likely capable of moving civilization forwards, so... :shrug:
Yes but also the robotics industry notoriously has an extremely low shipping rate. So choose your bog I guess?
And some of them are military focused.
Eugh. No. Let him create stuff and let someone else take care of the taxes and payroll and the sales and all that other stuff.
He might be perfectly happy doing projects and YouTube.
Likely makes more money, too!
For many, no; taking on lots of sponsorships, you can make a good amount of money (especially the ones where you agree to do X posts across Y social media accounts for Z period of time, essentially being fully sponsored across a large swath of life).
But for a lot of tech/engineering channels, it'd be immensely difficult to make the same salary as you could working at a FAANG or the like. (I'm making about half what I made when I had a W-2, but it's enough).
Do fun projects of your choosing vs the grind of corporate life.
The former has a rather large non-pecuniary component of total compensation.
Exactly; especially in terms of stress—which is a major problem with my chronic illness, and both the main reason I'm happy to be away from a W-2, but also the thing that makes me most nervous as I'm dependent on marketplace healthcare (which is quite expensive for much worse benefits), and it's harder to get things like life/disability insurance as an independent.
Being a top YouTube producer is not without stress : Reacting to the change in number of followers, responding to comment that can be unpolite, managing sponsors...
If you’re not in software the salary you’d be making as an employee would be much lower than what you refer to as “FAANG” though
Thank you for taking a pay cut to do what you do, I'm so glad your channels exist. Also please remember to take vacations! (trade shows don't count)
You just need to start selling supplements, anti-aging creams and time shares.
His current Patreon has 57 paid members. Unless he has another revenue stream, I think he'd need more paid members.
He is a senior mechanical engineering student at purdue. I think the channel is something he is just getting rolling in his spare time.
youtube ads and sponsorship are his additional revenue streams
[dead]
Very impressive. Commenting to be able to find this later because I need to keep tabs on this guy. The CD launcher is incredibly cool.
Nobody hire him!
Another cool project from this person's blog: "Wrist Mounted CD Launcher" -> https://www.aaedmusa.com/projects/wrist-mounted-cd-launcher
It definitely gives off Elysium (film) vibes.
I think he'll have success with youtube/vlogging more than getting into the corporate world honestly - especially with some healthy sponsorships and great projects like this.
What is the power consumption of these robots? I often wonder how limited and viable autonomous robots really are. When I look at Tesla's Optimus or Boston Dynamics' spectacular robots, how quickly do they need to be recharged?
My understanding is that those motors he uses are pretty special. And I would expect “efficient” to be part of that special. You’re optimizing for torque and accuracy per unit of mass and energy in this sort of space. I know he talked about them in earlier videos but I no longer recall the details.
Aside from the onboard computing required to power the AI model controlling the robot, which also consumes energy, it is important to be as efficient as possible. However, if an Optimus is intended to replace human labour, such as lifting objects and wander around with it, the energy efficiency of the platform becomes negligible.
The platform in the video and the robot dogs from Boston Dynamics are ideal for tasks where they are only limited by their own weight and the amount of computing power required to navigate, such as exploration.
I suppose that's why we mostly see autonomous delivery robots on wheels.
Or maybe I am being too pessimistic about other platforms...
This project is so cool! Congrats Aaed.
Breaking Taps on YouTube did a really awesome video on a somewhat similar mechanism (I'm no mechanical engineer haha, it was new to me!), rolling contact joints. I love the idea of using string/ropes. Worth checking out as well if this kind of stuff interests you! https://www.youtube.com/watch?v=TQiLLcumqDw
Some resources related capstan drive: https://www.osti.gov/servlets/purl/1340266 https://fab.cba.mit.edu/classes/865.21/topics/mechanical_des...
Wow! I actually met Aead last week while he was printing parts for this project (we both work at the same place). Surreal to see it at the top of HN.
I had never seen a rope use a robot dog before so I clicked immediately. It wasn't what I expected from the headline, but very cool nonetheless.
This is an absolute masterclass in DIY robotics. With a proper battery and a tougher foot material, this thing could really stomp.
Speaking about battery, how much current does it draw while at rest?
Why spend so much effort to achieve an "exact" gear ratio? Having more zeros does not equal to being more "precise".
Also, I wonder how resistant this mechanism is to wear and fatigue.
Well, it probably wasn't that much effort. When you're 3D printing you're going to end up printing everything 2-3 times anyway, so why not dial in the ratio while you're at it?
And you can't really declare your design is "high precision" and present yourself as someone others should take transmission design advice from if you aimed for a gear ratio of 8 and achieved "somewhere around 7.9 to 8.2"
It probably doesn't matter so much whether it's 7.913 or 8.186, but it would be important to know the exact value for kinematics. One way to do that is to build an object very accurately, the other is to build inaccurately and then measure the result after the fact.
It's also interesting because competing actuators with strain-wave, cycloidal, or planetary gearboxes will state exactly what the ratio is. The actual gear teeth may not be spaced out perfectly around the circumference, but the number of teeth is an integer with an infinite number of zeros.
Yeah, I think one of the nice things about making it a "clean" number (either an integer or a rational with a small integer denominator) is that you can easily validate it without needing high-precision measurement equipment: put a mark on both gears (maybe even embedded in the 3D print), line up the marks, rotate the large gear 1 full rotation, and count the number of rotations the smaller gear makes. Check to see if the marks line up perfectly after those rotations.
His capstan reduction can't go all the way around even once.
It could though. I don’t think the thought has occurred to him yet. He could make the stack twice as high and go around twice as far. If you moved the worm gear you could go farther, but I don’t know how he would do that with his drive.
He could also go with narrower rope, and spread the load over more windings, which would give him more throw.
What are the loads on a drive in this use case? 35lbs robot bouncing on one foot would be 350lbs-ish of dynamic load? That can be handled with 2mm Dyneema.
Getting that ratio nailed down also makes future designs more predictable I think
> Also, I wonder how resistant this mechanism is to wear and fatigue.
He actually discussed this in an earlier video for his initial tests on the capstan drive. He ended up testing the rope he used for around 358 hours (two weeks) on continuous use in the drive itself with very low backlash
https://www.aaedmusa.com/projects/capstandrive
https://youtube.com/watch?v=MwIBTbumd1Q&t=10m
Because when a real engineer puts 2 and 2 in and gets 3.8 out, it vexes them and they want to at least know why they can’t get 4. He’s trying to make a machine that does what he told it to do, so that he understands what is actually happening.
I think it's about kinematics, the more precise your gears the better the model fits the real world.
That's why pro crews don't use gears and ropes. At high impulses deformations and elasticity throw the kinematics off what's actually happening. Modeling the deformations and the elasticity is a computational no no. Instead what you see is the motors right on the joints.
At least that was the case last time I had a look at robotics.
> That's why pro crews don't use gears and ropes. [...] Instead what you see is the motors right on the joints.
The answer here, as with so many things in robotics, is: It Depends.
UR10e robot arm that can lift a 4kg object with a reach of 1m and has sub-1mm repeatability? Strain wave gears in the base and shoulder joints, 100:1 ratio.
MIT Mini Cheetah robot dog that can do backflips? 6:1 planetary gearbox.
Shadow Hand with 20 degrees of freedom? Tendon driven, with the 20 motors in the forearm to keep the fingers slim.
Little dinky Huggingface SO-101? Servo motors, integrating 1:345 gearing with a series of 6 tiny brass gears.
Mid-price CNC milling machine, if you call that a robot? Really long ballscrews, driven by stepper motors.
Surely you mean driven by closed-loop servos with encoders, don't you? Jacques Mattheij wrote a long post about how he ended up having to replace all the CNC machines he sold that used stepper motors because he didn't know any better at the time, and brushless motors are a lot faster and more powerful as well as not losing steps.
I guess "mid price" isn't the most helpful term when machines vary from $100 to $100,000 huh?
I think the X-Carve 3-axis wood carver uses stepper motors with belts of all things. The Shapeoko Pro is leadscrews and stepper motors. Wazers, I believe, are belt-and-servo driven. And a lot of 'CNC conversion kits' you can order online use stepper motors. Plus of course laser cutters have really low torque requirements, so they've got a lot of design freedom.
Arguably those are "cheap" rather than "mid price" it just felt weird to declare a $4000 machine to be "cheap"
Actually, it was the Shapeoko 1 which introduces the idea of belts (MXL), then switching to Gates GT2 belts with the SO2 (ob. discl., I got a free machine for doing the instructions, see Github). Since it was opensource, the X-Carve was forked from the SO2, though the X-Carve Pro was a completely new design.
Note that belts were continued through the Shapeoko 3 (I got a machine as a "thank you"), though the Z-axis got a leadscrew in the Z-Plus upgrade, then the Pro (since, I got a job with the company and got an XXL as part of my employment), then the 4 (the original Pro is referred to as a 4 Pro sometimes), and it is only with the 5 pro that ball screws were switched to for all axes (and I now have a 5 Pro).
For an example of what a belt-drive CNC can do see:
https://community.carbide3d.com/t/hardcore-aluminum-milling-...
I think there are CNC machines that cost 100 times more than US$100k.
Some products should be compared on a log scale. CNC machines are a very good example.
The middle of a log scale from US$100 to US$10M would be US$30k.
In general, for some platforms each gear mechanism adds backlash precision loss, lower energy efficiency, and might not be back driven.
>Mid-price CNC milling machine
A ball-screw is mostly decorative on small machines... =3
On the other hand, a 100:1 gearbox gives you much-needed torque if you're lifting a load at the end of a long arm, it makes your encoder 100x more precise (in terms of repeatability) and it makes your motor brake 100x stronger.
Back-drivability is the enemy of precision, so many robotic applications can do without it.
Almost all modern servo driven units I've seen prefer to allow some compliance in the end effector. The UR5 and UR10 series for example can use force limiting control loops, and are safer to use around people.
The old "fast, cheap, or good... choose any two joke is mostly still true. =3
The UR10 uses 100:1 strain wave gearing in the base and shoulder joints.
You’re right that it has a freedrive mode, and force control modes. But it’s a rigid, low-backlash robot with the compliance achieved in software afterwards.
Expensive, naturally, but none of the problems that come with things like series elastic actuators.
> At high impulses deformations and elasticity throw the kinematics
Sure. Yet evolution has achieved astonishing kinematics with all manner of deformation and elasticity inherent to the materials, and also constantly changing physical properties, using low resolution data. We cannot build permanently lash free mechanical devices at reasonable cost and reasonable size/weight. Eventually, the answer must be pervasive real-time compensation throughout the kinematic model.
> Modeling the deformations and the elasticity is a computational no no.
Why? Nervous systems do this. That's why you can change your shoes and still walk upright.
In the video he mentions how the specific type of Dyneema cord he's using is well-suited for the application (compared to other kinds of rope/cord). It's particularly strong, light, and inelastic; a lot of climbing equipment uses a version of it for similar reasons.
more than 30 years ago I was writing code for a "robotic" device that used motors, directly on the joints.
the motors were so sloppy the company wasted a ton of money [0] having me write heuristics to tackle the errors they accumulated over several hours.
one of his whole points is that by using dyneema (rope), there's almost no elasticity at all in the capstans.
[0] relative to the cost of better motors
> Modeling the deformations and the elasticity is a computational no no. Instead what you see is the motors right on the joints.
That sounds like “It’s not wrong, we just don’t do it”. There are some amazing examples of imprecise drive systems compensated for by excellent control systems all over the world, for millions of years.
>> Having more zeros does not equal to being more "precise"
Isn't having more decimal places the exact definition of precision (vs accuracy)?
The point is that those decimal places don't have to be zeros. 7.893 is just as precise as 8.001.
Yet for some reason unknown to me, people get annoyed when you tell them „let us meet at 13:37. that is no more accurate than let us meet at 14:00 hours”
14:00 is a Schelling point
Even small deviations can compound over time in a real-time system
That was confusing part of this video . May be there are some limitation on the tools he uses to tune
I don't think the number of the gear ratio really matters, what matters is that you know what it actually is (since every IK calc depends on said ratio); 8:1 is probably arbitrary and/or looks nice & might simplify some stuff.
It might be a lot easier to check the ratio "by hand" (by counting rotations etc) if it's numerically simple. (IIRC in some earlier videos he noticed that the pulley size ratio wasn't producing the expected movement ratio, because they were built as an obvious 8:1 or 10:1 or something, and didn't match - which led to him figuring out the subtleties of the design - I can easily imagine wanting to preserve that aspect just for debugging, at that point, even if you now have correct math.)
From a coding point of view it’s also nice if all the drives are exactly the same, so each isn’t compensated for separately. But yeah, just a nicety.
Thanks for sharing this! What a treat of a video. It's a fun project, and it's presented very well. This guy has a talent for communication - the video was super clear and well explained. I really admire that ability and I want to get better at it.
Cara is the Irish word for 'friend'. Not sure if thats what was intended or is just a coincidence.
Male genital organ in few of the balkan languages (spelled with K), pronounced the same. So, there's that too.
Also the Spanish words for "face" (no me gusta tu cara, boludo) and "expensive and female" (esa computadora es re cara, boludo). https://en.wiktionary.org/wiki/cara tells us it also has meanings in Aragonese, Asturian, Catalan, Crimean Tatar, French, Galician, Indonesian, Italian, Latvian, and another couple dozen languages.
Just saw this at Open Sauce last weekend! Super cool project, looking at all the gearing was fascinating.
This personal site makes me genuinely jealous—in the best way.Really awesome side projects and great intro.
Time to raise my own bar.
Professor of Upstairs Neighboring. https://youtu.be/8s9TjRz01fo?t=1128
There is a weird mixture of hope and dread in me as I watch this. I am ridiculously excited over a person like me being able to mess around with something that, until recently, was gated behind, well, a lot of hard problems to overcome ( I am only now slowly getting through old Peter Scott's robotics to get some perspective ). By comparison, it should be so much easier to explore aspects of robotics that may go beyond strict math and engineering.
Gated how? It's all on textbooks, industry standards, Internet, everywhere. You could buy a KHR-3 without even driver's license for $1.5k cash for past ~20 years. I guess CUDA was technically gated behind "I swear I'm not a bad guy" button, but that was it.
I'm surprised that Ukraine isn't DIYing full-on solid motor cluster SAMs and armor piercing ATGMs. I thought those kinds of devices were something just about any sufficiently developed country can do in days to weeks should such national emergency arises and all bets were off, except nations in peacetime has moral obligations to do no evil.
Why dread?
Land mines have been killing people by surprise for decades. Flying drones, some of them dropping land mines and others lying in wait for incoming convoys, produced 70% of the casualties on both sides in the Ukrainian invasion last year, and a few weeks ago the Ukrainians shipped some drones to near a Russian airport several thousand kilometers from Ukraine to blow up a bunch of airplanes. This betokens a future of borderless warfare in which probably most of your family members will be killed over the next few decades, by one or another type of semi-autonomous or fully autonomous weapon.
CARA is a super cool project which is never going to kill anybody, but it's another piece of evidence that the cost of the technology for such weapons has decreased enormously.
That said, talking about the dread is going to get boring fast, because nearly every story on the HN front page is catapulting us toward that future.
Good point. If we do press panic button each time, the word will lose its meaning. Naturally, it does not help that this particular community did happen to catapult us towards the future a fair bit over the past years.
I'm more optimistic. I think filling battlefields with autonomous weapons will make humans obsolete there and reduce war causalities. Basically any attack on humans is going to be a war crime and terrorism akin to attacks on civilians russia is doing right now
Well, suppose you have two countries at war, say, Burnia and Hevonia. They each produce a million autonomous weapons.
Burnia sends their million autonomous weapons to kill the million people they judge are most crucial to Hevonia's war effort, prioritizing Hevonia's political leadership and military officers.
Hevonia, meanwhile, sends their million autonomous weapons to destroy Burnia's autonomous weapons, when they can find them, but not to attack any humans.
Who wins the war?
I think Burnia does, because even if Hevonia's weapons are 99% effective, Hevonia's government has still lost its top ten thousand people, including all of their military officers, while Burnia has only lost half a billion dollars. That's going to make it impossible for Hevonia to keep fighting. And I think 50% effective is more likely.
It's the same question as when one opponent is ready to nuke enemy cities while the other is not. In theory everything turns out the same way you postulate. In practice cost of winning the war in this manner might be higher then the cost of losing localized, limited conflict.
That is a really weird take considering how hard Russia (and Ukraine) work to continue to kill humans in the same situation you are describing.
Who cares if it's weird? Why are you saying "weird" as if that were a bad thing? It's just a derogatory synonym for "unusual". A take would have to be weird to be a contribution to the conversation, because if it's not weird, it's a widely held opinion we all already know, and nobody gains anything by reading it.
What matters, given that it meets the minimal bar of weirdness to be potentially worthwhile, is whether it's correct. Which depends on what future combatants will do once weaponry actually is autonomous, not what Ukraine and Russia are doing with remotely-piloted FPV drones.
I'd call it delusional than weird. US is clearly controlling Ukrainian churn rate so that Russian military and Ukraine as nation burn down at the same rate into disappearance so that the postwar Ukraine can be rebuilt with more Western leaning meatbags from surrounding nations. And some of people here are framing exactly that as utopian bloodless fights of machines. Calling it weird is itself almost weird.
I think we are in transition period and nations are just learning that sending people to contact line is a terrible idea. People never learned things like that very fast. It often takes a generation.
> This betokens a future of borderless warfare in which probably most of your family members will be killed over the next few decades, by one or another type of semi-autonomous or fully autonomous weapon.
You're talking about a whole bunch of targeted and intentional attacks in a literal warzone, the most lawless kind of place on the planet and you're complaining that they are hitting their intended military targets?
Meanwhile Gaza is boring to you, because Israel uses conventional bombs with humans in the loop, even though the collateral damage matches your prediction today.
I'm not complaining, exactly; I'd rather the Ukrainians be blowing up Russian planes, which were thousands of kilometers from the literal warzone, than the Russians blowing up Ukrainian planes thousands of kilometers away.
I'm saying that the events we're seeing in the Ukrainian war are convincing evidence that the nature of war has changed, and the implications of that change for human society are disquieting.
For two million years, people have often resolved conflicts by warfare. But that warfare, though it has never ended completely, has always been localized, which meant that most places most of the time were peaceful.
This is the foundation of the system of international relations in which different states exercised monopolies of legitimate violence over geographical territories: by so doing, they could prevent warfare and provide security. This system has existed to one extent or another for thousands of years.
Warfare is no longer localized, so now the only place for anywhere to be peaceful will be for everywhere to be peaceful. It is no longer possible for states to provide security to the people within their borders. Consequently statehood itself has lost its meaning. The last time we saw such a change in the nature of warfare was twelve thousand years ago when the first states arose.
I have some corrections for you about other things that you did not understand either about my comment or about the wars we are commenting on.
It bears repeating that the planes blown up in Operation Spiderweb were not in a literal warzone, nor anywhere close to a literal warzone. Some of them were thousands of kilometers from the literal warzone, in a part of Russia that has China and Mongolia between it and Ukraine.
The Ukrainians reportedly did have humans in the loop for that attack in particular, and both the Ukrainians and the Russians are mostly using FPV drones rather than autonomous drones of any sort at this point, unless we count landmines as "autonomous drones". In https://www.youtube.com/watch?v=RmfNUM2CbbM Russian drone developer Sergey Tovkach specifically says that Starlink is the reason Ukrainian sea drones have established control of the Black Sea — Russia could build sea drones but does not have an equivalent to Starlink, so evidently both Ukraine and Russia are dependent on remote piloting for their sea drones. He also says Starlink's latency is too high for piloting quadcopters. (Evidently Operation Spiderweb used the Russian cellphone network.)
The Israeli attacks are also using precision-guided weaponry and (in some sense) AI, not conventional bombs, and they're killing lots of innocent people, but you're right that it's less worrisome to me, for three reasons:
- The total number of people they're killing is about an order of magnitude smaller, about 80,000 so far (out of a total of 2.1 million Gazans), versus roughly a million in the invasion of Ukraine. Many more civilians have been killed in Gaza, but both Russia and Ukraine practice conscription, so it's not as if the soldiers being killed are only or even mostly volunteers; they were civilians until being, in many cases, conscripted.
- The weapons Israel are using are mostly very expensive, which limits the number of people they can afford to kill with them. By contrast, the drones being mass-produced in both Ukraine and Russia cost only a few hundred dollars each, and each country is expected to produce about 3 million of them this year.
- The people Israel are killing are almost entirely in close geographical proximity to Israel. Gaza City is 80 km from Jerusalem. Beirut, the residence of many of the Hizbullah personnel that Israel killed with explosive charges in pagers (along with, in several cases, their children), is 230 km from Jerusalem.
Therefore, while Israel's war is clearly causing terrible suffering to millions of innocent people, and it could easily spiral into a Third World War, it does not represent the advent of a new, borderless mode of warfare in which the cost of untraceably killing a precision-targeted human anywhere in the world is similar to the cost of a small air conditioner or 20 kg of beef.
The developments we are seeing of the mode of war in the Ukrainian invasion do represent such a transition, and that is true regardless of how the conflict goes. But mostly it is not the Ukrainians and Russians who are responsible for the transition; it was an inevitable result of developments in batteries, motors, power electronics, and 3-D printing and (more broadly) digital fabrication.
What will the new equilibrium look like, now that statehood is effectively meaningless? Well, it might be better or it might be worse, but not many people currently alive will live long enough to find out.
Fascinating! I want to get into this type of stuff. But I have no idea where to start, I just have just a CS degree and 3 years experience as a developer.
I recommend a Brachiograph build. It will introduce you to some fundamentals of PWM and inverse kinematics. It is well documented but not cookie-cutter. Using a Raspberry Pi will give you more direct access to running the servos than the microcontroller experience. All the parts are infinitely reusable afterward if you don't want to keep it around.
https://www.youtube.com/watch?v=u4Jh1daCl60
https://www.brachiograph.art/
https://github.com/evildmp/BrachioGraph
A solid place to start is building something simple
Arduinos and hobby servos. No, neither of them are "industrial grade" and yeah, you'll reach their limits pretty quickly, but building a physical thing that does stuff is (in my experience) a huge motivator.
Or if you're already all over the basics, figure out what kind of stuff you want to build and then try and build it. :)
Can you recommend where to find beginners projects that “do stuff”.
To begin with you can get a cheap robotic kit from Amazon, there are many of them, and put it together. That's probably the easiest and fastest way. From here you can read more about servos and controllers, modify its mechanics and software.
Just like the article, there's a huge amount of hobbyist-accessible projects on youtube, you can click around the recommended videos.
Grab a servo and start playing with it.
https://hackaday.com/ - cool projects and interesting stuff https://news.sparkfun.com/tags/tutorial - tutorials on embedded software, electronics, etc. https://learn.adafruit.com/ - electronics tutorials
New here and just stumbled upon this. Seen these "robodogs" live in Vegas mining conference. The usage is picking up, their functionality and range of movement is so much more advanced now, and the list of actions they can complete much more complex. Awesome stuff!
Very cool robot dog and interesting video! Can the dog climb stairs? Isn't capstan drive temperature sensitive, e.g. the ropes will be shorter in cold and longer in warm wheather?
Nice to see a cuadruped robot that doesn't sound like a swarm of angry bees.
Similar mechanism might be good for astronomical telescope mounts. Gears are often a problem and nobody seems to have solved it.
I remember a post previously related to him here on HN but I am surprised at how I forgot about him and how cool he is to build all these incredible stuff and also teaches in his videos. I am subscribing so that I never miss anything from him again.
What a horrible video is this, with the robotic translated AI voiceover?
Update: Ah, weird, if I watch the non-embedded one on youtube it is the original in English with normal sound. It's the one embedded on his web site which has AI translation to German.
That's not the video's fault, but YouTube's. YouTube has started adding automatic AI voiceover translations to all videos, unless the uploader explicitly disables them for each single video. It's opt-out and the option is well-hidden, so most uploaders will not disable it.
As an end user, there's nothing you can do to prevent from seeing them. But you can change the audio track to the original while the video is playing.
> YouTube has started adding automatic AI voiceover translations to all videos, unless the uploader explicitly disables them for each single video.
WHAT?!? It a time when there is a growing animus towards AI generated content, for many content creators this would be tantamount to slander
https://github.com/YouG-o/YouTube_No_Translation
How do we make such robots intelligent? Like you only tell it to learn to jump the rope and then it goes through trial and error(s) to figure out a way to do it, and if it can’t and needs a physical upgrade then it tells you that? (Like a certain gear ratio etc)
https://developer.nvidia.com/blog/introducing-isaac-gym-rl-f...
It's just a gut feeling but I'd trust a feedback based backlash elimination mechanism more than rope based, especially in the long run and/or with a large deployed base.
But, why?
Like I said gut feeling. I suppose over time there may be some play in the axles, for example. Something feedback could compensate for but a rope could not.
Anyway good engineering requires extensive testing over many hours of movements.
> Anyway good engineering requires extensive testing over many hours of movements.
He actually discussed this in an earlier video for his initial tests on the capstan drive. He ended up testing the rope he used for around 358 hours (two weeks) on continuous use in the drive itself with very low backlash
https://www.aaedmusa.com/projects/capstandrive
https://youtube.com/watch?v=MwIBTbumd1Q&t=10m
The problem with gear is that they suffer from latency.
Amazing work, beautifully presented.
Geordi LaForge reminds me of Aaed.
I recently found these videos. Amazing guy, amazing skills, great humor.
He’s got the magic combination of tech skill plus ability to make and edit entertaining videos.
incredibly impressed everytime i watch one of his videos
> Programming takes the cake for what is both my most and least favourite part of any build. Nothing quite makes you pull out your hair and ask yourself, 'What the heck have I gotten into?' Like spending weeks programming a robot that just won't work. Eventually though, you fix that one line of code that makes all the difference. And then it's smooth sailing. Well, kind of.
I feel this deeply, also this whole video is quality content.
this guy's work is pretty amazing I whole heartedly believe that I'll own a gang of (armed) quadrapets for home/property/personal protection within the next 10 years. I'm equally worried that once these become commodities they might be worse than an electric scooters have become to us bipeds on sidewalks but on balance I can't wait
This is absolutely amazing, awe inspiring.
I'm so excited for a smaller open source version of this, I'd love to make one. What a great project.
It might be fun to optimize the shape of the rolling contact surface of the capstan drive away from cycloid to make it even better suited for specific application, like a robot dog leg that smoothly runs for miles.
[dead]
[flagged]
I don't think it is fair to downvote this comment. It is a genuine concern and should be addressed. Amusingly, given that the thought has entered mass consciousness by means of question on 'whether all this is a simulation', matrix ( the animated series ) explored this question a little and it is interesting in how the timeline aligns with what the movie presented.
> It is a genuine concern and should be addressed.
No disagreement, but does the comment meaningfully contribute to the discussion about this particular project?
Fair point upon reflection.
It's absolutely a legitimate concern but since this robot project has no chance of replacing humanity, it seems like a derail in terms of the subject matter of the video. It's as if someone posted a link to an incremental advancement in graphene aluminum-ion battery tech and a person chimed in with a general lament about heavy metal battery toxic waste disposal.
That's an interesting reference, but do our timelines really align with what was presented in (ani)matrix? Their timelines are quite murky, but it seems that in their universe things only started getting problematic between humans and AIs around 2090 and The First Machine War started in 2139 [0]. With how fast things are progressing, it's not clear to me that we have so much time.
[0] https://screenrant.com/matrix-timeline-year-future-when-sett...
Definitely.
It's intellectually exciting.
On the flip side, capitalistic / private / special interests both controlling the progress and having the most ability to utilize it to further centralize power and wealth is deeply concerning. We can already see more controversial figures involved in AI using it to spread their personal viewpoints.
It feels really easy to see how our jobs/labor and therefore our capital and therefore our value in the modern system are being directly attacked by these capabilities and deeply concerning to imagine how further centralization of power could be good for the masses.
I don't like these robot dogs.
See also the youtube video about the project: https://youtu.be/8s9TjRz01fo
The linked page has that video embedded.
Yeah good point!
When I came across this amazing project and wanted to share it to HN, I was debating whether to post the youtube link or the project page. I decided to post the project page and mention the youtube link in the description for those who prefer video, but somehow that description got posted as a comment instead (not sure how that happened?). Anyway as you said the video is embedded in the project page so it wasn't really necessary