I really dislike this writing structure of "before I get to the point, let me tell you a story about my life"<p>Edit: I'm not alone, <a href="https://style.mla.org/dont-bury-the-lede/" rel="nofollow">https://style.mla.org/dont-bury-the-lede/</a>
I’m “unlocking” that this will be dead within 18 months. It will get too hard and bored engineers will head back to Google Corp where the promotion game is better.<p>I’m in manufacturing. Machinery is highly specialised. Making a generic robot without taking up huge amount of floor space and/or huge leaps in programming is like… Kubernetes being good for hosting your moms book club blog.
Makes sense. The end state appears to be that humans should only be supervising ML that generates goal and outcome based behaviors for robots, and the machines will construct tools to solve problems themselves.<p>The leap from an AI model learning how to replicate a behaviour (e.g. evolving walking to solve problems <a href="https://unitylist.com/p/2id/walking-ai" rel="nofollow">https://unitylist.com/p/2id/walking-ai</a> ) to reasoning about it in terms of actuators and physical feedback, to assembling a physical model out of a relatively small list of parts seems like a solvable engineering problem when it is broken out into a pipeline.<p>Those robot parts are basically a version of mechano with actuators that a model would map a behavior to, and the robots in the article would assemble them. When you look at something like Lego or Mechano as an intermediate representation to construct buildings out of, where all objects made from it are essentially a directed graph of those elements, robots designing and building robots seems like less than 20 years away.<p>e.g. we could functionally specify to an ML model, "produce a digraph of these element parts that has these degrees of freedom, and then load or derive a model that solves for this outcome within the domain of those degrees, where outcome is 'plug cables into a board' "
Nice.<p>Here's much the same job, being done almost 50 years ago, by a robot at the Stanford AI lab.[1] This robot has both vision and force feedback, and uses them to assemble an automotive water pump. It does the coarse alignment visually, and the fine alignment by feel.<p>[1] <a href="https://archive.org/details/sailfilm_pump" rel="nofollow">https://archive.org/details/sailfilm_pump</a>
Billion dollar question: do you get more bang for the buck (return on investment, ROI) out of improving robot control schemes, or out of designing the product with automated assembly in mind?<p>Bonus: the ROI changes as you invest in either bucket.
Funnily, their scroll capture is totally broken on their website [1] at least on my version of Firefox. Also, it's strange that they don't have their own website when they're a separate company. I guess it shows the ephemeral nature of these projects.<p>[1] <a href="https://x.company/projects/intrinsic/" rel="nofollow">https://x.company/projects/intrinsic/</a>
Off topic: Wow, X's main website[1] is infuriating. The scrolling is very janky with a touchpad, and the carousel at the bottom of the page which highlights projects in a timeline advances way too fast to read it comfortably, and there isn't any obvious way to pause on a slide. Shame, because I was actually interested in learning more about their projects.<p>[1]: <a href="https://x.company" rel="nofollow">https://x.company</a>
I used to work a bit with industrial robots and I found the biggest issue wasn't a lack of innovative technology but how hard it was to access: Everything was an expensive "option", the programming languages and tools were dinosaurs, and documentation was treated like some kind of sacred text, only for the eyes of the exalted.<p>Maybe with a "software" approach to this we'll see better and more open tools.
If you can't program it by directly showing it what to do, throw it in the bin. They probably intend for it to be SaaS too, effectively having you pay a workers wage for a truly terrible worker. You're competing with general intelligence robots that cost $12 - $15/hr. That's 10 years of full time labor for $300,000. No shot.<p>This looks like something designed to attract ignorant investors/talent who think small time manufacturing looks like a Ford plant but with less robots and more humans. In reality it looks something closer to Grandma's kitchen on Thanksgiving. How are you gonna stick a robot in there and have Uncle Fred program it?<p>I can't see this as anything other than a flashy high school engineering project. Much wow! little application.<p>Source: Work in domestic manufacturing. <$50 million company. Mostly do government/military electronics building.
Sensing and control are certainly part of the problem, but to me it always felt like a major limit to automation was the quality of actuators. It's much more than just a control problem to make robot hands with the sensitivity, acuity, and dexterity required to crack an egg, thread a needle, and play Chopin.
It seem that this company is doing more of the middleware and higher level interfaces/adding intelligence to industrial robots than they are trying to build their own robots (Google tried that at least 3 times and failed).<p>Anecdotally, I've heard that FANUCs don't respond well at all to any input deviation.
This is great. A concerted effort to make robots be able to assemble and make more things.<p>This is a piece of the puzzle of building a machine of machines that can make almost anything without human intervention.<p>Are they hiring interns?
I notice a lot of pessimism -- presumably at some point someone is going to build much better robots than the ones we have today. (Surely we are not anywhere close to the end game of robotics). I'm inclined to believe Google, with all of its resources, has as good a shot as any.
> Back in the late 90s when I was just starting Moonfruit, the world’s first SAAS website builder, creating your own website was hard. From setting up your own server, to working with an ISP, to getting a content delivery network and integrating a middleware layer to communicate with your computer, to design and UX — creating a website was a lengthy multi-step process that was only accessible to a small group of technical experts or large companies. It wasn’t until websites were simple and easy to make that the full creative and business potential of the web really began to blossom.<p>It's not good that this introductory post doesn't start right off with a problem to be solved. Instead it presents the credentials of the current leader.<p>If I had to pick out the problem, it would be this sentence, contained in the fourth paragraph:<p>> Currently just 10 countries manufacture 70% of the world’s goods.<p>In the fifth paragraph, we get a more clear phrasing of the problem:<p>> The surprisingly manual and bespoke process of teaching robots how to do things, which hasn’t changed much over the last few decades, is currently a cap on their potential to help more businesses.<p>Ok, so this is going to be a company that solves the problem of poor usability of industrial robots through machine learning. The larger goal is to put manufacturing capacity closer to consumers for better sustainability.
>...the US manufacturing industry alone is expected to have 2.1 million unfilled jobs by 2030.<p>Is the implication here that they're aiming to automate away all of these jobs?
Well they don't have intrinsic.com or twitter.com/intrinsic... Those are still associated with a tech startup from a couple years ago:<p><a href="https://twitter.com/intrinsic/status/1164007322932277249?s=19" rel="nofollow">https://twitter.com/intrinsic/status/1164007322932277249?s=1...</a>
Over the last few years, our team has been exploring how to give industrial robots the ability to sense, learn, and automatically make adjustments as they’re completing tasks, so they work in a wider range of settings and applications. Working in collaboration with teams across Alphabet, and with our partners in real-world manufacturing settings, we’ve been testing software that uses techniques like automated perception, deep learning, reinforcement learning, motion planning, simulation, and force control.
> "when I was just starting Moonfruit, the world’s first SAAS website builder"<p>Moonfruit, launched in 2000, was definitely not the first SaaS website builder. Geocities launched 6 years before it and there were dozens of them by the time Moonfruit came around.<p>While not a big lie, it's an odd way to start a post like this.
TL;DR: making industrial (e.g. manufacturing) robots easier to use, by improving sensing, planning, etc.<p>I suspect that Dr. Chelsea Finn's work in meta-learning (affiliated with Stanford and GBrain, when I saw it last year) might play a big part here, which is e.a. about generalisation of RL policies to out of domain tasks. (E.g. similar task, but slightly different tools, slightly different task, etc.)<p>Learning IRL (cameras and actuators) reinforcement learning policies is a huge time sink, so generalisation is a hugely important task. Related solutions can be found in simulation->real generalisation, also an active topic of research.
Robotics is not a software problem and SV companies bias is towards software development (a little different with X but still apparent). I think most companies that try to throw data at existing problems in robotics using existing machines will have a hard time matching human efficiency. For example, in something as straightforward as the usb insertion task.<p>Hardware and mechanical is like 95% of the problem so there's a need to take the approach of making the machines that make and then add the software on top and developing synthetic task orientated data from that. E.g. the dishwasher, which works because its physically designed for washing plates and then automation was added. The robot arm is a general purpose technology that has been around in the same form since the 60s/70s. There are many options as alternatives (e.g. magnetic assembly or even self-assembly in certain industries) but ofc these are incredibly risky commercially.<p>I'm aware that this is just the first post and the above is well known in robotics development so excited to see what gets built!
So they are scripting Kuka robots effectively?<p>Well, actually if they do some AI stuff that might be impressive.<p>I guess stationary robots are seen as less of a reputational risk in comparison with Boston Robotics nighmares.
To me, this is <i>robot vs process</i> - how much do we need clever robots and how much do we need to change the job.<p>There is an old saw about the transition from steam powered factories to electrical power. Initially the large steam engine was in one location, and basically its power was delivered by belts running off one central location. The factories initially tried to replace the steam engine with one big electric motor, and it worked ok but the factory was still a hub and spoke and pieces had to be moved from one spoke to the next.<p>It was not until a new generation of factories were built with <i>many</i> motors at any point in the factory that the modern line was built.<p>Of course this is a massive simplification, but I look at two robots using 10 m2 to assemble some Ikea cabinet, and think "awesome geekery" but if you want a factory producing pre-made furniture go back at least three-steps.<p>Robots that can replace a human arm in the assembly process just feel like we are replacing that big steam engine in the middle of the factory.<p>And, yes industrial robots is where you start, of course. But a factory can change its process to eliminate the need for a general purpose robot. But <i>the home</i> - that's a different story.<p>* Take up two "normal" sizes of a washing machine. A hopper accepts clothes, sorts them using RFID tags, and begins a run in a smaller drum, spins, dries and folds them. (yes, its probably magic but this would be on everyone's XMAS list)<p>* (completely foregoing everything I just said) a mobile robot arm that can learn where each item in a house belongs. 3D tracking, ML etc, and it picks up the toys my kids have left lying around.<p>* I am not sure where the "robot" vs "process" sits here, but food purchase and prep is a large time sink for many, but there seems to be a viable disintermediation of supermarkets - I mean if i choose a decent set of meals for a week, why send the food to the supermarket so it can use its shelves as a collection point to send it on to me. And if the food is picked so i get "nice meal on Saturday" plus "something with the extra Tues lunch"<p>I think there is a real possibility of robots making the middle class home like a B&B.<p>As Jerry Hall said, "My Mother told me if I wanted to keep a man I needed to be a Chef in the Kitchen, a Maid in the living room and a Whore in the bedroom. I said I would hire the first two and take care of the rest myself."<p>Edit: honestly I am not trying to be HN-negative, and I think all this investment is only going to build better robots. Which is a win. But I remain under-convinced that building general-purpose robots to replace general-purpose humans, when humans are already having the easy bits replaced by specific purpose robots is a good idea - it feels like running uphill.
The domain name reminds me of x.com, Elon Musk's 1999 company that became PayPal. It was one of only 3 single letter .com domains. I have a memory that its issuance was a mistake or some sort of strange deal but I can't find any evidence for that now.
The x.company website is unusable on Firefox. One scroll wheel movement and I am lost on a completely different part of the document. There is one way to ensure an immediate bounce.