Hacker News Reader

Comment by btbuildem | Hacker News Reader

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.

View on HN

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.

  > It does make me wonder about 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 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)

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.

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 am interested. I love these builder channels.

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 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.

"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”.

Interesting you've mentioned sailboats. I've encountered dyneema for the first time in spearfishing!

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.

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.

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.

The Sandia paper also claimed that steel cables need a larger pulley radius.

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.

These drives sound amazing, so why are they not used everywhere? What are the disadvantages?

Reply on news.ycombinator.com