I can understand why the style may seem offputting, but the thing to understand is that it has been traditionally very hard to engage with the public on this topic of robotic advancement. In fact, I know a bit about this myself, having been in the robotics space for over a decade. But my own struggles in the field only reflect a longer trend, which I can even trace back to my grandfather.
Growing up in a strict Lutheran household in the southwest England town of Flenkelshire, Elias Nathaniel "Kazoo" Pendleton III did not immediately stand out among his peers. Born with dull red hair, one leg three inches shorter than the other, and shoulders that somehow resembled cornish hens, young Elias was a frequent target for the town bullies. A child at that time has only three options: fight harder, run faster, or invent some kind of device that would enable him to escape his tormentors. Luckily (by chance or by fate), Flenkelshire was home to a radio-electronics store, Bundleron's Radio and Horseshoe Supplies, which gave young Elias just the right ingredients to hatch his escape plan. And hatch a plan he did, though it would take twenty years for the town to understand exactly what had happened.
The first trap was set in the Fall of 1951. Winston Churchill had returned to power. The Festival of Britain had just wrapped up and lit the imagination of attendees and non-attendees alike. And Elias Nathaniel "Kazoo" Pendleton III, now well-armed with a stock of electronics, metalwork, and several years of intense study, went into action...
SeriousEats. It’s truly amazing. Most of their recipes are split into two pages, one is just ingredients+steps, and the other is a “story” - but not an irrelevant story of a person, dish, or how it tastes on a warm summer day, but instead it lists different experiments the author tested, results, dispels common myths, …
That hits so close to home. I’ve given up searching for recipes online because of this. And I’m not even mentionning all the ads you have to scroll through.
It's viewed as a DRM measure. Recipes are not copyrightable unless they are attached to a story. It's probably an urban legend, but can't blame poor food bloggers from acting on it.
Based on this reasoning, the United States Copyright Office Compendium, the Office’s manual for examiners, states that a mere listing of ingredients or contents is not copyrightable, as lists are not protected by copyright law (chapter 314.4(F)). The Office has also stated that a “simple set of directions” is uncopyrightable.
In addition, courts have found that recipes are wholly factual and functional, and therefore uncopyrightable. As the Sixth Circuit described in Tomaydo-Tomahdo, LLC v. Vozary, “the list of ingredients is merely a factual statement, and as previously discussed, facts are not copyrightable. Furthermore, a recipe’s instructions, as functional directions, are statutorily excluded from copyright protection.”
I just think they're part of the non-fiction fantasy genre of entertainment alongside cooking, travel, and house buying + renovating shows.
They're ostensibly informational, but 99% of people consuming them aren't genuinely looking to cook the thing, travel to the place, or buy and renovate a house.
Not much extra work surely. Most recipes are on sites that use the same template for every recipe so it's pretty easy for a scraper to find the ingredients and method. The structure is usually even labelled with meaningful class names on the elements.
Yes - it's not specifically that the recipe sites have changed to do this, but rather, that Google is preferring to show you sites that are doing this.
I dislike this story in journalism and podcasts too. I listen to a lot of true crime podcasts:
> Before Y was murdered, they lived in X. X is a quiet town, the type of place where you don't need to lock your doors. Y has a happy upbringing collecting flowers along the river at...
Like we get it, this is the first half of every 1-hour long true-crime podcast. Also quite often the first half of every long-form article.
I don't mind it as much in true crime podcasts when it's done well. Totally agree that the generic "it was a peaceful town where nobody locked their doors blah blah blah" can get old quickly. But hearing about the unique lives of the victims in murder cases can definitely add to the story. And in podcasts more focused on the investigative side (ex. Someone Knows Something) knowing the background info can even be critical to solving the puzzle so to speak.
Oh gosh I hate this about wondery business wars[0]. Very informative but also very annoying when they conjure up complete/narrated conversations between people e.g. a tech CEO and an investor or customer. This happens a LOT in the show.
Maybe I'm off base here, as it's no small feat by any means, but it's especially jarring to read when "first-website-building-SaaS" has so little to do with industrial robotics at its surface.
I had a similar impression. It was jarring to say the least and made me ask so many questions that had nothing to do with the story. Does this person lead every tech related conversation with, "When I was the CEO of Moonbeans, the world's first SAAS blockchain beanbag chair crowd sourcing platform" just to buy some credibility? Why do they feel the need to tell us that? Do they have inadequacy issues or is it the opposite? I can barely remember what that article was about. Robotics? Oh, right, shame it's an Alphabet subsidiary. It's bound to end up in the Google graveyard when it fails to be one of the top 10 most profitable companies in the world. Even if they do great things and make a great product, it's the fate which will inevitably follow most of Alphabet's projects until the SEC breaks them up, this being one of many good reasons.
I like inverted pyramid style too, but it’s a very brief intro and letting the CEO of a new company introduce themselves doesn’t seem so bad? You could skip the first paragraph.
I hat this too. Especially for newsletters. They tease you with an interesting headline and then let you scan the entire thing to find maybe just a link to the topic teased.
When typesetting a printed newspaper, the leading (pronounced like “ledding”) is the space between lines of text, originally physical strips of lead (Pb).
The lede (“leed”) is the most important statement in the story. The word comes from “lead” (also pronounced “leed”), because it's the statement that everything else should follow. It's conventionally spelt differently to avoid confusion with “leading”.
Off-Topic: I would recommend aspiring writers to read, or better still, write, academic papers. Once you get the hang of what makes a good paper, it helps with all forms of non-fiction writing.
I agree. First two paragraphs are completely irrelevant and offer no support to the actual topic. Just flexing on career. Kinda sounded like typical LinkedIn content these days
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.
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.
I worked at Google X for two years on a different robotics effort.
In general you’re not wrong. It’s also an easy prediction to make because almost all new robotics efforts fail.
I’ll just say that they’ve been working with manufacturers for years on test projects. So this isn’t really a new thing so much as a formalization of whatever they came up with.
I’m still not sure how this stuff will go (X is of course for experimental stuff) but if the project has “graduated” to a company than I’d presume they have some revenue plan and potential customers. I could imagine that at least one of their trial projects with a manufacturer has worked out, and they think they can find more partners for this work. (I only scanned the beginning of the article)
They have been experimenting and developing a LOT of stuff since 2013 and plenty of it sadly does just get completely sacked and mothballed (and I have my personal gripes with the great projects they could have open sourced but instead they just let the bits rot). That said I think with how much I saw them working on they probably could find applications where profit is a real possibility.
Given that this is Google, my money is on them either abandoning it, or making a new version every 2-3 years (for no discernable reason) making no actual progress as a result.
I would LOVE to be proven wrong, but Google's track record with bringing products to market, and actually keeping them for longer than a media boost, is downright depressing.
At some point it feels like every Google side project is just a media buy for their real business: hiring engineers to drive more ad revenue, engineers who are enticed by shiny projects like these.
It’s not depressing, it’s wonderful. Google’s abject failure in almost every area is great. They have nearly unlimited resources and a fantastic (though but as great as they think) engineering pool. It would suck if they managed to successfully attack adjacencies.
I have a hard time calling immense wastes of resources "wonderful". Think of how many man-hours of at least better-than-average engineers have been wasted developing something like 5 separate chat apps... it's absurd.
Totally agree with that. There are ways to do it, but the approach they show is doomed to fail. Arrival had excellent ideas and kind of working products but if nobody injects money in them they are done pretty soon.
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.
The leap from an AI model learning how to replicate a behaviour (e.g. evolving walking to solve problems https://unitylist.com/p/2id/walking-ai ) 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.
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.
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' "
This is not manual or bespoke and it has sensors. The videos are incredible and they work in real life already.
This one is it moving petri dishes full of liquid without spilling! This is obviously not being pre-programmed to move along some kind of 1980s style fixed paths for welding parts as Alphabet apparently thinks everyone is still doing. The obliviousness of suggesting that using ML models for robotic control is some unique new idea is really off-putting. Mujin has been around since 2012.
The more the merrier, of course, but just dismissing the state of the industry and claiming you've made a huge technology leap (compared to the 80s and 90s instead of something harder)... ugh.
> as Alphabet apparently thinks everyone is still doing. The obliviousness of suggesting that using ML models for robotic control is some unique new idea is really off-putting.
Intrinsic/Alphabet are not suggesting they are somehow unaware of easily-Google-able state of the art in ML robotics. They literally used to own Boston Dynamics.
From the post, the second demo of their tech (“Two robots use perception, force control, and multi-robot planning to assemble a simple piece of furniture”), is very clearly much more than “moving Petri dishes”.
FAANG has access to the leading factories in Shenzhen, and heavily utilize robot tech in their HW supply chains.
It's pedantry. FAAG is not an acronym in common usage (and is uncomfortably close to being a slur), so the more easily understood, less correct word was used instead. To point out that one of the companies doesn't produce consumer hardware doesn't invalidate the underlying point, so what is it,
if not pedantry?
I’m pointing out that it’s stupid to use an acronym that includes mega companies that don’t have access to cutting edge Shenzhen manufacturing and excludes ones that have far more access.
Facebook doesn’t have better Shenzhen access than Microsoft or you know, hardware focused companies like nvidia.
Using “FAANG” is a red flag that the commenter has no idea wtf they are talking about when it comes to the hardware industry.
The few Mujin videos I watched look a lot like PCB assembly pick-n-place machines. A little bit of computer vision, a little sensing here and there, but overall fairly simple pre-programmed moves, on a pretty controlled environment.
If you check out the beginning of the video link (I had it fast forwarded towards the end) you can see that it is doing an awful lot more than that, and in 2013.
A pick and place is 2-axis movement with a suction cup. This is controlling a robot arm with a ton of degrees of freedom and developing paths for moving through all those degrees of freedom without hitting anything and using internal models to do so.
I suppose in some very broad sense it looks similar, but the difficulty of x-y + down is way, way lower than what you're seeing in that video.
It is harder than x-y + down, however I don't think this video is impressive really, having slowed down the video it doesn't look special to me and I did work on robotics/machine vision around that time.
Humans design products with a variety of design elements to meet different circumstances.
Look at cars for example. Tell an ML model to "make a car that can drive over rocks" and it will give you a rock crawler with the motor in a location where it won't be easy to fix. Tell the ML model to "make a car that is easy to fix" and it will make a car that is probably unreliable. Tell it to make a car that is reliable and easy to fix you will get a car with no motor at all.
I'm not saying it's impossible, because it obviously is possible. I just think your 20 year time-frame is hopelessly optimistic. What good is an ML model that takes 10 weeks to setup that solves a problem that only takes 2 weeks to solve without ML?
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.
You're just grumpy :-). On the plus side the computer that is controlling the robot isn't a DEC-10 in a climate controlled room so there is that.
The correct feeling here though should be compassion, here is a group that has been safely nestled in the arms of Google X and is now being pushed out of the nest like so many projects before it, which currently has one such company, Waymo, that is currently not yet dead. Statistically speaking, it is unlikely they will be able to pupate into a products company before they run out of time.
That said, it is also a truism that the constraints on robotics 50 years ago are not the constraints on robotics today. Re-implementing those ideas which had merit before but lacked a sufficiently robust ecosystem to be practical might in fact be really useful today. One hopes that they have the perspective of the excellent technical reports that SAIL produced to guide their development.
It's hard for me to compare precisely, especially since the Intrinsic videos are sped up, but the one you linked looks very shaky and hesitant, and also the "Ikea challenge" seems like it requires more fine-tuned force-feedback than putting metal pieces together. If I anthropomorphize, the Stanford robot looks like an inept/hungover employee, whereas the Intrinsic robot seems convincing that it's actually accurately aware of what's going on.
Another possible difference -- how much programming time did it take to teach the Stanford robot to assemble the water pump? Sounds like Intrinsic trained the robots to do this with little supervision.
It seems to me that this might represent pretty solid progress, although not exponential/paradigm-shift scale like we've seen in some other industries in that period, and nothing in the Intrinsic videos seemed like it was above par for other automation companies I've seen recently. But since you seem to be in the industry, what's your take on whether they seem to be ahead of the game, or even just realistic, with claims like:
> In one instance, we trained a robot in two hours to complete a USB connection task that would take hundreds of hours to program. In other tests, we orchestrated multiple robot arms to assemble an architectural installation and a simple piece of furniture. None of this is realistic or affordable to automate today — and there are millions of other examples like this in businesses around the world.
It's hard for me to compare precisely, especially since the Intrinsic videos are sped up
Here's a longer version in a larger size, either not sped up or not sped up so much.[1] It's using a simple strategy of approaching the socket at an slightly off angle and then twisting into alignment.
That's a standard strategy. Compare this video of assembling Lego blocks with an industrial robot. Note the little twist moves.[2]
Did the machine learning come up with that, or was it preprogrammed? Did ML re-invent remote center compliance? That would be progress.
Rod Brooks went down this road with Rethink Robotics. They went bust.
You can certainly do what they're showing. It's making a profit on it that's hard.
Thanks for the link. It seems to me that this is rather specialized, and would run into trouble with designs that have a very small base. Also, the number of parts is limited. Imho, this is closer to a printer than to a general model assembly robot.
https://youtu.be/cNxadbrN_aI here are neural networks in the 50s. Something having been tried before doesn't mean it isn't worth ever trying again in a context of much improved technology.
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?
Bonus: the ROI changes as you invest in either bucket.
Apple was once into design for assembly. The Macintosh IIci was Apple's peak at design for assembly. It was designed for vertical assembly. Everything clicks into place with a straight-down insertion move. No wiring harnesses. The power supply plugs into the motherboard. An automated plant in Fremont CA did the assembly.
Then Apple gave up on design for assembly and went to offshoring and cheap labor.
Motorola flip phones were designed for automated assembly. All parts were on boards, and the boards were stacked and compressed into a solid block, with bumps on the boards making connections to the next layer. A tough, reliable phone resulted.
Then Motorola gave up and went to offshoring and cheap labor.
Sony pioneered this approach. The Sony Walkman,, the original tape unit with motors and contra-rotating flywheels, was built for vertical assembly and assembled by a simple Cartesian robot.
Apple case is not exactly abandoning design for assembly. The advancement of electronic and metal machining allows smaller and more integrated parts, which allows cheap labor to beat the machine. If the electronics and metal machining did not advance, I am guessing the resultant production cost would not be this low.
I don’t think is true at all. CAM is dramatically more advanced than it was in the 70s - easier to use, and better algorithms mean much faster pathing. Costs are way down. Tooling is cheaper and more reliable.
Of course there's been quantitative improvements, but fundamentally everything that can be designed and machined today could be designed and machined in late 1970s using very similar tools, processes and control systems.
Yes, the technology is the same. But how many machined products were available for the consumer to buy in 1970? I was excited about the Macbook Air not because it was thin, but because it was CNC made, just like the aerospace products I designed. Injection molding remains dominant, but over the last 15 years CNC has made a lot of progress.
I dunno, by the end of 1970s most high end camera lenses were CNC machined with bezels and scales CNC engraved. My 1980 Summicron 35 is certainly one, although the optical stack sits in an injection molded envelope.
When it comes to MBA enclosure it's likely the infill or separation problems with long thin walls that led them to abandon injection molding. Even then it's not done on a mill in one take. Things like speaker grilles are probably something like EDM or etching rather than 8000 operations with 0.2mm drill bit.
This was absolutely machinable with late 1970s CNC mill. Probably pain in the ass to design with then existing CAD/CAM but ultimately this demo runs on same G-Code that was there in 1970s.
And as far as parlor ticks go this is not particularly impressive, compared to old ones like turning cube inside a cube on a lathe.
That’s a bit like saying “most large modern systems are written in C which was invented in 1972, therefore software engineering has not changed since the 70s/80s”. That machine is also probably not running a gcode flavour that would have made sense in the 70s anyways, with additions for active sensing, macros, etc.
Tools like UNISURF would probably not be able to handle the extreme level of detail on most of that model. The very long tool lengths you see in the video are much more complex than they look and require effort both from software and hardware to prevent chatter and breakage on titanium. The clearances in the video are also extremely tight in places, and while you could have guess-and-checked that in the 70s, it’s a very different workflow than the simulators that are basically standard use today.
CNC machining is a multi-billion dollar industry populated by smart people. While the fundamental technology of “spinning cutter driven by computer controlled motors” hasn’t changed in 50 years, the R&D departments aren’t asleep at the wheel.
Machining grade 5 titanium alloy (the most common one) is not especially challenging; it's easier than 316. Saying this as I did machine it personally on a manual lathe.
From the initial endmill size they use and the generous machine sizes (which can be extrapolated from chuck/table size) we can easily tell rigidity/chatter wasn't a particular challenge here. And the finishing passes are so delicate they don't even bother with coolant for that.
Yes the code to run all this would be a monumental undertaking in the 70s but it is done once and can be still entirely cost efficient as far as mass production is concerned. The thing new machines have is the running speed (for same precision work). These did improve a lot and make certain classes of products economically viable.
> Do you, or anyone else here, know how much money Apple saved by offshoring?
There's another axis here, which is how our desire for a product overlaps with DFM. It could be the case that offshoring to cheap labor actually increased the manufacturing costs 2x, but enabled a product that would sell 10x better than its DFM counterpart.
(I have no data to say that is the case, only the intution that these things are complicated systems which rarely come down to single-issue decisions.)
Half of automating tomato harvesting was breeding a tomato that could survive the harvesters. (Granted they also taste relatively horrible, but now we all eat them because they are cheap and ubiquitous.)
The improvement of robot control schemes reduces the constraints on designs that are designed for automated assembly. I suspect there’s a kind of slow moving coevoultion there where you have to go incrementally at both.
Indeed. Looking at their sample footage of assembling Ikea furniture reminds me of fixturing. Watch manufacturing footage and watch out for jigs and fixtures. They are EVERYWHERE.
Currently, you can either use fixtures and jigs and specialized machines and run fast, use humans and run medium speed, or use AI and generic robotics and run REALLY SLOWLY.
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.
Wow, you're right. At least on Linux with Firefox 90.0.2 scrolling with the mouse wheel moves the page by such huge steps that it's basically unusable. Had to navigate with the arrow keys instead.
Right, they seem like a project not a company. I suppose they can put in a request to have Alphabet fix the website, they may not have the autonomy for that themselves.
Everyone thinks that they are somehow different, but all these firms fail for the same reason. Robotics is hard. The market is not that big. Lots of costs. Investors are skittish. The combination of those things isn't that good.
Your conclusions are true, though I think it may be helpful to further subdivide into some categories as far as what the target market was and where they were at with technical readiness.
Like, some of them (Anki, Jibo, Mayfield, Asimo, Reach) were 100% toys, and were always going to be at the extreme end price-wise trying to compete with increasingly "smart" toys being manufactured by regular toy companies with regular toy company processes, volumes, and margins.
Others (Rethink, Willow, Schaft, Blue) were trying to do something really ambitious and potentially provide B2B value, but were never far enough along to have a compelling value proposition for the end users they were targeting. They were never fast enough or reliable enough to be competitive with the minimum wage labour that they would have displaced— if robots are hard, then mobile robots are harder, and mobile manipulators are the hardest of all.
I think the saddest story in here is still Starsky, because they weren't in either of these groups: they really did have a clear value proposition, and they were technically there as far as delivering on it. The market needs what they were offering; they seemingly just ran out of runway at a time when investors were too starry-eyed about vaporous promises of L4 autonomy to want to back a company working on a viable hybrid solution.
(Disclosure: I work for a B2B mobile robotics company)
>They were never fast enough or reliable enough to be competitive with the minimum wage labour that they would have displaced
This probably sums up well. Human are extremely adaptable. To point if we are measured as 100 then no Robot is even 1.
There is a whole reason why even Foxconn gave up using Foxconn Robot, some task are just insanely easier and cheaper for a human to do it. They’re not easily automatable and even if we could the cost benefits doesn't make any sense.
So instead of having human plugging in DIMM RAM or M.2 SSD, now they are all soldered on the logic board using machines with automation.
That may be true today, but it might not last forever. Labor cost in 1st world nations is skyrocketing (due to cost of living mainly) compared to poorer nations, and there may come a time when robotics becomes relatively competitive. Especially when those cheap labor countries start having the same effect.
I think a lot of that comes in the form of partial and adaptive automation, though— like self-checkout at the grocery store, where it's "automation", but only in the sense that the self-checkout console enabled outsourcing the pick and place part of the work onto the consumer.
Or elsewhere in the thread, the example of moving a previously-modular computer part onto the logic board, so that it can be soldered on rather than needing to be installed later in the assembly process.
Companies like Rethink weren't in this world— they were trying to build a manipulator (Baxter) which was a drop-in replacement for a person doing pick and place work. Which has a certain appeal, if it works ("no need to retool anything; just buy it and put it to work!"), but it puts you up against the direct price comparison of just having a human continue to do that job.
Well, while cost are high in first world country, labours are mostly limited to services sector.
In manufacturing most of these labour are still in Asia. And the cost / productivity is still insanely cheap. It isn't just the cost of the Robot itself, but to program a new task which requires software testing and engineers. So the cost barrier is still so far apart. Foxconn make hundreds of millions of smartphone every year. You would have thought saving $10 per phone would have net them a few billions extra profits. And yet their employment rate has remained largely the same.
If and If, US and Tech managed to do this ( there is nothing even remotely close in the next 10 years, but let say somehow there is for the sake of argument ), this will be the largest reset of manufacturing and likely be Industrial Revolution 3.0.
> program a new task which requires software testing and engineers
To be fair, Rethink did understand this part, and part of their pitch was that it was supposed to be easy to teach their robots tasks with a kind of observe/repeat flow; here's a video from way back in 2012 showing where they were trying to go with this: https://www.youtube.com/watch?v=gXOkWuSCkRI
They're not the only ones either, UR has also placed a heavy emphasis on safety and ease of task training, though unlike Rethink, I don't believe their systems come with any built-in sensing, so it really is limited to just mindlessly repeating exactly what you show it: https://www.universal-robots.com/academy/
> Especially when those cheap labor countries start having the same effect.
You missed the comment’s pivotal point. As developing countries, well, develop, higher labor prices will affect the entire supply chain. It’s a Good Thing (TM), and that’s why we’ll need better robotics in that future.
Starsky value prop was teleop, but that was the same thing that cooled investors. Adding an extra 20-100ms latency to driving is akin to driving after two drinks. Operating a vehicle 10x larger than the ones on the road does not make this problem smaller.
Operating large trucks is not a game VCs wanted to play.
The point was that it was an autonomous system that could ask for help, and the "help" scenarios would mostly be cases where the truck was already stopped or at very low speeds: navigating a construction zone, a transfer yard, etc. Possibly in some of these situations it wasn't even wheel-to-wheel, but rather a system of choosing between a handful of high-level courses of action for the machine to then proceed with, or helping the perception system classify an unknown object it was looking at.
I didn't sense from the postmortem articles by Stefan that safety concerns were what killed it. It was investors being disappointed that they weren't trying to build a truck without a steering wheel at all, since that was clearly where Uber, Waymo, Tesla, and others were headed (and at least at the time, external safety concerns were not seemingly impacting any of them).
I just don't think you can call that a real value prop if it's only for when the truck is stuck or a few minor edge cases. There are many scenarios where self-driving may not work or behave erratically so if their version of teleop doesn't solve those then not sure how Starsky argued they were ahead of competition.
Additionally I think investors backed out primarily because of risks associated with operating an autonomous fleet, not the shortcomings of the tech itself.
I feel that it covers an awful lot of them. If you cap teleop driving at 20km/h or something (or maybe a dynamic cap based on your rtt), that still covers all of the parking lot scenarios, as well many sensor-failure situations, like if you needed to crawl along in the right hand lane because it's a blizzard and the radar is blind.
In any case, the Forbes article specifically addresses how they modeled these things:
"Up ahead a deer jumps into the truck’s lane and hundreds of miles away a teleoperator is asked to take control of the vehicle. But they aren’t able to in time – either the deer jumped too quickly or the teleoperator wasn’t able to get situationally aware or worse yet: the cellular connectivity isn’t good enough!
Such was the situation painted to me time after time after time as CEO of Starsky Robotics, whose remote-assisted autonomous trucks were supposed to face exactly such a scenario. And yet, it was an entirely false scenario.
As I’ve written about before, safety doesn’t mean that everything always works perfectly, in fact it’s quite the opposite. To make a system safe is to intimately understand where, when, and how it will break and making sure that those failures are acceptable."
The fleet argument also confuses me; hasn't that been the Waymo/Uber pitch since forever, a centrally owned and managed fleet of autonomous vehicles for hire? Why would that be considered an especially risky direction?
> We also saw that investors really didn’t like the business model of being the operator, and that our heavy investment into safety didn’t translate for investors.
This is what Stefan said here [0]. Honestly I hear contradicting reasons for the failure. It could be that their investors had a different risk tolerance than Waymo/Uber's.
I guess I'm confused, sure, teleop could cover a lot of the edge cases but if there is a fat long tail you still end up with a pretty unsafe technology. The deer example is kind of a distraction and goes to show that maybe Starsky had a problem imagining and classifying catastrophic failure events. For every deer jumping in front of the vehicle there is a 10x more serious scenario that could lead to human fatalities.
After reading his posts I'm still confused about the reasons they failed. Can you list the reasons from high priority to low as to why they failed?
The issue with most pure-robotics-that-make-things[0] companies is that they end up finding out that they need to iterate on the robot while the actual product gets better. It's not like software where essentially everyone can use the same spreadsheet. It's "oh, I need this panel here to have a 3mm smaller gap" which works when you're Tesla, because the product is the company, but it doesn't really work when you're just trying to make a series of robots that solve generalizable problems. Reality isn't as standardized as a Turing tape. Too many dimensions, figurative or literal.
[0] As opposed to robots that, say, fight wars. But we call those things "missiles" and "fighter jets" and "drones" not robots.
Missile robots are the best: you don't really need to worry much about supporting legacy products years after selling them to customers, and they are not expected to be functioning after just one use.
> you don't really need to worry much about supporting legacy products years after selling them to customers
You really, really do. Missiles are expensive, and stay in inventories for a very long time, and they need to be made compatible with every update to every platform that can make use of them. That wouldn't be so bad, but then you also need to prove that they work with all those platforms. This is hard.
> they are not expected to be functioning after just one use.
Missiles are only fired once, but that doesn't mean they are used once. The typical "use" of an aircraft carried missile is that it is attached to a plane, powered up, and then the plane does a sortie and lands, and then the missile is removed and maintained. There is a lot of maintenance that is done to the missile daily.
Market is not that big? What is the size of transportation industry alone? What about ride hailing? Investors are skittish? Cruise raised $10B most of it not that long ago, EmbarkTrucks is merging with a SPAC to go IPO soon, I could list others. Robotics is hard but that’s kind of the point.
I agree with this analysis, although I’d disagree that it’s questionable, I’d say it’s straight fallacious. I’d draw a parallel to the development of CNC technology[1], in the case that if this software solution can become successful, it seems feasible to me that their might become some sort of equivalent to a machine shop, but for assembly/robotics instead of manufacturing/machining. Currently we have Foxconn, who is doing significant research in the manufacturing automation space, and seems to be making progress, but I see no reason this couldn’t take a similar arc. CNC/CAD was initially only for the most ambition prototypes, but as it proliferated it reshaped the product market, making curves easy and allowing for much more complex 3d shapes, and was kick started by the stagflation of the 70’s.
I don’t look forward to (more) products put together by machines that are impossible for a human to do. But I genuinely feel that mastering robotics is one of the most important goals for society as a whole (and especially for safety conscious western countries), up their with clean energy and carbon sequestration. There is a lot of manual labor that (especially) Americans need to do, from updating infrastructure for rising seas and fixing the poorly maintained infrastructure we have, to increasing housing in urban centers, to whatever form carbon sequestration ends up taking––and western disease leaves these countries mostly unfit for the task ahead.
Transportation might be the identifiable target market, but the actual market of buyers for robotics in transportation is very small, and the problem is that the chasm between the incumbent market and new entrant robotics space is far too large to surpass by the emerging startups.
Amazon bought Kiva a while back now to do robotics for them, and it's used heavily in their warehouses and facilities around retail side. Anything they can automate through robotics, they try to as robots can work 24x7 (other than maintenance requirements) and over their life span cost less than human workers. They also sponsor engineering competitions around trying to make generalised picking machines. It's good PR for them, and although unlikely any time soon, someone _might_ have an inspired idea and solve something that has vexed experts for a long time.
I think the key here is that Intrinsic is (apparently) focused on designing new interfaces for existing, proven industrial robot models, rather than being focused on novel hardware R&D (a monumental task).
The head of robotics there mentioned that it was a strategic move because OpenAI wanted to focus on AGI. If OpenAI had other goals, like improving robotics, the division would still be around.
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.
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.
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.
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?
I can't see this as anything other than a flashy high school engineering project. Much wow! little application.
Source: Work in domestic manufacturing. <$50 million company. Mostly do government/military electronics building.
Assuming 8hr work days and a 250 day working year for humans. No such constraints for robots. That's 2,000 hrs/yr for a human vs 8,760 (assuming 24/7) hours for a robot. Obviously there will be other costs for robots (downtime, electricity, repair etc) so no telling whether it will be worth it in the long run but the hour calculation there does seem a little off.
When you are working on domestic manufacturing, what automation you have? How much of that is programmed by human (in house and from vendor?)
This is definitely far from mass adoption. But somewhere certain expensive product might benefit from this. Guess: mechanical watch assembly, given the amount of manual labor, and the claimed learning ability, it seems possible for a robot to assemble a 1milions worth of Swiss watch.
These are complementary developments, Boston Dynamics is building robots that excel at navigating the world in the way its built. This seems to be intent on building robots that excel at interacting with the world in the way it wants.
The latter has a much broader customer base. From picking fruits to folding shirts to installing a headliner into a new vehicle, there are many applications.
A bunch of their public projects incorporate robotics in some way, just from glancing at x.company/projects. Everyday robot, mineral, wing, loon, waymo, makani. I'm sure there is a lot more going on, like intrinsic.
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.
I agree, and the cost. Automation hardware is just fundamentally really really expensive. I guess part of that is due to the small market, but I'm a bit skeptical that they will ever bring robot arms to the masses just because robot arms are super expensive.
The ability to plug in cables and whatnot looks like a useful ability but I'm guessing this will just be sort of like really good traditional robotic control software rather than anything really fundamentally different.
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).
Anecdotally, I've heard that FANUCs don't respond well at all to any input deviation.
That's similar to the way you could say the founder's startup Moonfruit which launched back in 2000 was more about producing Flash SWF files than websites. Maybe the model has some traction for a group of people who find that sort of help applicable or not harmful to their own business model.
I currently have a robot washing my clothes, another robot washing my dishes, a third robot vacuuming my floors. Cybernetic systems adjust the temperature in my house and water my lawn. I even have a little robot I keep in the freezer to make ice cubes while I sleep.
So, two answers - 1. folding laundry is a difficult technical challenge. 2. when we get a robot to do that task, we won't call it a robot.
If it has an arm sort of thing doing it, we very well may. If your drying machine had an arm putting clothing up on pins in your backyard vs being a box you dump clothing into, we’d likely call it a robot.
Some people do refer to automatic vacuums and other things that move automatically robots too.
My quick test for a robot would be: does it have at least one appendage-like part performing a task? A washing machine has nothing like that, but a Kuka robot practically is a programmable arm, and Boston Dynamics robots have legs.
Folding laundry (and more generally, manipulating deformable objects) is actually a pretty tough task for robots. So there has been some research on it, and you can find videos of robots doing it (very slowly). I guess we'll get there eventually, but right now even if it's possible, it's not at a level of speed, robustness and cost where it makes sense.
Because folding laundry is extremely difficult for robots, easy for humans and people are not willing to pay much money or give up much space for a laundry folding robot.
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.
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.
If I had to pick out the problem, it would be this sentence, contained in the fourth paragraph:
> Currently just 10 countries manufacture 70% of the world’s goods.
In the fifth paragraph, we get a more clear phrasing of the problem:
> 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.
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 purpose of the first paragraph is not to present the credentials of the leader. The purpose is to make a parallel between the current state of the robotics industry and the creative & commercial expansion of the web once the technology became more accessible.
The implication is that noone wants to work those jobs anymore and the options are to either import illegal workers being paid below minimal wage or replace by robots.
Or that Americans aren't willing to work for the prevailing global wage that manufacturing workers demand. Why would any company keep manufacturing here when labor is so expensive? Effective manufacturing robots would allow manufacturing to move back to the us.
How could this possibly work for, e.g., electronics? How does the government buy computers that are made in America? Unless there is some loophole whereby all the parts are acquired from wherever they are acquired, and the manufacturer just assembles the box in the US and gets to label it "Made in USA"?
I'm guessing this comment got down voted because Americans are against immigration in general. But whether its a good idea or not the fact remains that our desirability as a place to live could be used to satisfy any labor shortage.
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"
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.
While not a big lie, it's an odd way to start a post like this.
Funnily enough, one of those dozens of examples hits close to home here:
A main source of the original fortune that funded the creation of YC and thus Hacker News was the $49m sale to Yahoo! of Viaweb, a SaaS website builder (focused on ecommerce) founded by Paul Graham, Trevor Blackwell, and Robert Morris in 1995.
TL;DR: making industrial (e.g. manufacturing) robots easier to use, by improving sensing, planning, etc.
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.)
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.
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.
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!
But it is a software problem. Surprised that you mentioned robotic arm, which is basically just 3-5 servo/stepper motors connected to case and not super complicated to build with 3d printers. It's the software that powers it. Boston dynamics robots are not the top of the line in terms of hardware. It is the software that gives their robot the power to even stand up, which anyone who has coded the robot knows it only looks easy.
There is a software authoring problem (which is where the ML bits are crucial).
If we had to program all robots like we had to with CNC machines, then programming them would be a high skill problem, even if we throw a lot of tools at it.
I can work my way through a Tormach, but is that really what I want to spend time with? The ultra low level specification of what I need done?
I'd love a pedal based training system with something like "Identify", "Orient", "Place", "Count", "Test" to teach it things in steps & get a program out of the demonstration (that donut computer vision project was amazing, because it showed you didn't really need ML to do these things).
Like we have people who are demonstration learners, I wish I could do something like that of going from many scenarios to a final one and have the robot to dissect every one of my actions into a flow-chart of its own.
Sure software is crucial to the final working of the robot and it's not solely all in the physical design. Robots are not possible without software but I think the fundamental problem in robotics for manufacture is about physical intelligence and industrial design and engineering.
My approach would be to manufacture custom arms for particular tasks and in principle 3d printing the arms is exactly what i'm getting at (e.g. that optimised physical design processes save on cost and improves performance much more than software + expensive externally manufactured arms). 3D printed arms with comparable repeat accuracy would be an excellent optimisation over buying v expensive Kuka products. Then you could start think about different mechanisms (compliant mech, soft parts etc) and control systems/software.
Kukas are not really just a couple of servos (e.g. encoders) and there are many examples from the 90s of self walking robots with little software too. There's good literature on "morphological computation" or Rolf Pfeifer's book How the Body Shapes the Way We Think: A New View of Intelligence.
I'd say it's very different with X. I looked at the large number of hardware design docs that were open-sourced[1] when they shut down Makani - hell, even the Makani documentary[2] was mostly about hardware (material science, mechanics, aerodynamics) with a some software sprinkled in.
To me, this is robot vs process - how much do we need clever robots and how much do we need to change the job.
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.
It was not until a new generation of factories were built with many motors at any point in the factory that the modern line was built.
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.
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.
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 the home - that's a different story.
* 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)
* (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.
* 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"
I think there is a real possibility of robots making the middle class home like a B&B.
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."
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.
Lots of grocery delivery services do use purpose built warehouses. Stores like Walmart aren't doing that because it would cost a bunch extra vs picking from the stores they already have.
The furniture assembly thing probably doesn't make sense for huge runs, but you could stick one in front of a modest warehouse and build 200 different products on demand.
But I don't see the business case. Take the furniture - Ikea can assemble their flat pack furniture for me ... in front of the store. That's exactly where I don't want it assembled. In fact we are back to the value of the robot being in the home not the factory.
On suoermarkets, yes Walmart and Tesco made a sensible decision to use their existing stores as feeders for delivery. But as the number of people taking food deliveries goes up, that starts being a commercial disadvantage - there are things you can do to improve a warehouse that you can't do if customers are walking around in it, there are car parks that aren't needed now, smaller stores in expensive locations. It's not like Walmart's going bust next week but the world is changing. Even Amazon will feel this - a fleet of international transit and drivers designed to drop one package off randomly is going to find that model under threat if I get a food delivery every three days - 99% of the time I will just have the other stuff i order come in the same box.
For every house, it makes sense that only one delivery company visits that house. If they are delivering every couple of days anyway just roll it in together. Place tour order and it will be with your regular Tuesday drop off - that's convenient mentally for most people, whereas "expect a knock on the door anytime in the next 36 hours" is less easy to manage - especially now we are not all locked indoors.
I think i have wondered off the point but it's always much simpler to change the process than to build a human-analogue robot.
We all have huge shifts in how we will live in the coming decades - to stave off climate change, to take advantage of what software can really offer. Making already efficient factories more efficient is the equivalent of looking for your keys under the lamppost- the big wins lie elsewhere.
Edit - the big five future changes
- urban planning (see Tokyo, or StrongTowns)
- Buikding energy efficiency (solar panels, heat exchangers)
- Transoortation (Food and people) & transportation (other)
- ?
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.
Does anyone actually like these homepages that move around a bunch as you scroll? I don't see the appeal of parallax effects and the like. To my eye, they are neither pretty nor useful.
Same here. One detent of the scroll wheel and it's 2/3 of the way down the page. I tried disabling smooth scrolling, but that didn't fix it.
Note that we're talking about the home page, x.company, not the page linked.
Took me a bit to get to the web site, since I was lost in Medium land for a while, but there's an unlinked text-only "www.x.company" at the bottom I could copy/paste.
> I’m now leading Intrinsic, a new Alphabet company
> After five and a half years developing our technology at X, we’re now ready to become an independent Alphabet company, leaving the moonshot factory’s rapid prototyping environment
If you don't know what "X" in Alphabet/Google is, then this announcement looks a little vague and confusing, not like the skunkworks unveil it is supposed to be.
I think Hacker News tries hard to be constructive in its criticism and even upbeat. Consider the guideline: "Be kind. Don't be snarky. Have curious conversation; don't cross-examine. Please don't fulminate. Please don't sneer, including at the rest of the community." https://news.ycombinator.com/newsguidelines.html
One of the reasons I visit hacker news is because of the collective will to try to abide by guidelines like these.
It's the most effective Trojan horse if they actually succeed in general robotics.
Currently the amount of people using Google home is tiny. But imagine if you can just say "Google, make me a sandwich" and a few moments later the robot comes along.... And the automatic cleanup etc. People would buy them instantly, giving Google all the private data of households which would have never bought any of the current "smart assistants"
I'm pretty sure that household robotics is going to be the by far most profitable field if general robotics is ever achieved, as literally all households will want one, just like a washing machine.
The problem with household robotics is that people who have the dough for the maid-bot probably own stuff way too expensive to have it cleaned by a robot.
The Lusty Automaton Maid, on the other hand, is sure to sell like botcakes.
Edit: I'm not alone, https://style.mla.org/dont-bury-the-lede/