I invented Roomba and assure you, robots won’t take over the world
nautil.us
nautil.us
Automation will eat everything, it's just a matter of time. Even his problem of 'visually identifying weeds and uprooting them' does not come across as a massive problem, but one that can probably be solved with enough time and demand.
Platitudes like these give false hopes to workers who should be training and preparing for the next phase of industry, and IMO serves people like the author who need complacent workers to give them time to come up with their money-making solutions.
One thing I didn't quite realized needed teaching until I taught my daughter: how do you tell when a screw is going in right? It might have resistance because of gunk or rust, or it might be misaligned. Determining the right action is incredibly sophisticated, the difference between the slight easing of resistance with old oil or the springiness when the threads are pushing against each other and deforming. In response you might draw out and try again, or push harder, all while still trying to feel out what's really happening. Automation is great for screws in factories.
We don't have anything on the horizon that has the physical flexibility and power of a human. To the degree we can reconfigure problems to match a robot's capabilities, then yes, automation may be inevitable. A lot of problems aren't like that.
but i'm skeptical; it would be more convincing if GP/GGP provided some of the rationale for why middle managers would be automated away rather than simply referencing a book.
those kinds of jobs would have to be coercive in nature (desperation qualifies), and would be beyond the threshold of autonomy that most people will tolerate. most jobs, even those rigidly regulated, have enough latitude that workers can find some level of esteem in it, enough to create the intrinsic motivation to keep doing it.
people wouldn't tolerate such jobs without extreme coercion, implying nothing less than a collapse of our democratic republic to achieve.
> Platitudes like these give false hopes to workers who should be training and preparing for the next phase of industry
These seem incompatible to me. If automation will eat "everything," what exactly is the next phase of industry that we should all be preparing for? What's left after everything has been eaten?
My resolution to this conundrum is: automation will not actually eat everything. It's not like automation is a new concept; we started automating things over 100 years ago, and contrary to fears back then and along the way, automation drove tremendous job creation. There are far more human beings with well-paying jobs (even in May 2020) than there were 100 hundred years ago.
Billions of dollars are invested annually in technologies with marginal productivity improvements, and it looks like there are more and more diminishing returns.
If there was any indication that automation would 'eat everything', unemployment would be high (corona-pandemic aside), and labour productivity would be growing fast, none of which is happening.
Robert Gordon gives an expansive, evidence-based account of the development of American economic development in The Rise and Fall of American Growth: The U.S. Standard of Living Since the Civil War, and there's virtually no evidence for the grandiose claims made by the tech industry that we're living in some sort of era of rapid change. That for the most part ended in the late 70s.
Automation - handles deterministic situations.
Autonomy - handles uncertain situations.
In my industry (ship design, construction and operation), based on what I'm seeing and have seen evolve over the past 20 years, it feels likely that this paradigm shift to a majority of ships being autonomous will probably happen in the next 50 years. This is my gut feel, not a scientific conclusion, so take that for what it's worth. I also don't see why other industries will be any different. For my money, reinforcement learning (in some form) appears to be the key enabler.
I'm not sure what you mean by 'circlular', but I think it's safe to assume you mean a circular hull cross section vice a round circular deck (like a floating helo pad or SpaceX launch barge).
First, there have been some examples of surface ships that transition to submarines (e.g. 1). China is apparently working on a large Surface ship/submarine for their Navy.
Surface and submersible vessels have very different hydrodynamic and hydrostatic properties and forces to contend with, which drive them to have different shapes. I actually did work on a recent project that explored this topic a bit. Generally, for any vessel, the shape depends on what you want the ship to do and hull form/type selection is very much a science-backed art form of Naval Architects.
You can certainly build a ship to fully submerge, the question is, particularly at larger scales, is it worth the added cost, risk, and hydrodynamic compromise - or are you better off just building two different vessels that are specifically designed to operate in one domain (I.e. on the surface or submerged). That answer is highly dependant on what you intend to use the vessel for. It may be worth it for your application, but in general the answer has been that it's more cost-effective to just optimize vessel design for one operational domain rather that two.
Regarding hydrodynamic trade, while both ships and subs strive to minimize total hydro resistance (drag), hydro resistance has subcomponents that affect vessels differently when operating on the surface versus submerged. Surface ship shapes strive to be hydrodynamically 'optimized' to minimize form, skin, wave (critically), and even air resistance, whereas submarines aren't generally subject to wave action or wind and thus typically focus on minimizing the dominant form and skin drag forces. Surface ships are generally more power efficient than submarines at equal displacement and speed through water.
If you want to make a ship submerge, you also need to add significant structure to withstand hydrostatic forces at depth. Submarines are inherently structurally sound enough to operate at the surface, but function hydrodynamically poor at the surface because of their shape.
Surface ships and submarine shapes primarily differ in structure due to the significantly different pressure forces they need to contend with. Submarines have a circular cross section because, next to a sphere, cylindrical hulls are the most structurally stable pressure hulls. Surface ships can actually have relatively structurally 'weak' main decks, operating almost like floating buckets that are mostly designed to stay right-side up. The weaker main deck makes the ships cheaper to build, but probably more critically, allows easy access to whatever payloads the vessel is carrying.
1. https://www.thedrive.com/the-war-zone/9726/the-usmc-is-inter...
In general it's measured something like increase in $/employee.
But let's say you're a commercial printer. The industry is moving to new printers that print double the amount, per employee hour.
If prices stayed the same, you'd see double productivity.
But because every company has the new, fastest machine , prices greatly reduce, and stop reflecting the fact that a single employee can now do 2x the work.
this phenomenon (price deflation) can exist but we can actually measure deflation/inflation, so there's no need for a hypothetical, and there is no widespread deflation, or put differently increase in consumer purchasing power due to increased production.
There's actually very limited sectors where this dynamic exists (think the costs of a GB of data), but healthcare, construction, education, infrastructure all have not been automated, and hence aren't any cheaper.
In fact there's actually a decline in producitivity in US construction, as decades of erosion and retiring of experienced workers leave the industry in a bad shape, the US has started to lose manufacturing knowledge.
Whatsapp messages and tweets may have gotten a lot cheaper, but it's worth remembering that 'tech' is only 10% of the US economy, employing 3-6 million people depending on how you count. The overwhelming majority of the physical economy and service sector is still practically untouched by it, despite many claims to the contrary.
Just physically you can walk into a retail store, which is the most common place of employment around the world. It's got a Square Stand on the desk maybe, but otherwise, not exactly a lot of things have changed. Same thing in the classroom or a hospital ward.
That's true. But it's a bit more complicated.
In construction for example, there's the new prefab sector, which uses a lot of robots, growing nicely. It's a bit challenging to open such expensive factories with boom/bust cycles in construction, but they might figure it out.
As for infrastructure, we're starting to see companies working on robotic heavy equipment.
As for healthcare, that's an industry I don't understand. It's a very innovative industry. And some of the treatments and the medtech should have created efficiency gains, while not looking like robots. For example: curing hepatitis or cnc for dental crowns. But we don't seem to measure that efficiency. Weird.
As for retail, with e-commerce, some of our transactions have surely became more automated. And we chose to convert some if the efficiency gains, into comfort, to receive things at our door. But delivery robots might automate that part too.
I think it's inevitable that the "junior" office jobs (where many get their start) will be/are replaced by scripts & API's.
like fraud detection. before, you just accepted a 3% fraud rate as a cost of doing business and left it at that. now you get that down to 0.5% because the tools let you do it in the same amount of time, and the owners want that sweet, sweet cash. you still want a second accountant to double-check the first, so you expand the work to fulfill that need. it can still be worth it, even so.
Yes and no, people are predicting that for 70 years, if not more. It seems that once you automated all the easy things it gets exponentially harder and more expensive to pursue automation.
> does not come across as a massive problem
Sounds like Musk telling you "self driving cars will be there in 6 months" every 6 months since 2013. Nothing is as easy as you think it is, even more so if you have 0 experience in the domain.
And then you have to solve all the societal issues. If everything truly gets automated at some point, who will benefit from it? how do you remunerate all the people who used to have low skill jobs? &c.
"Automation, which is both the most advanced sector of modern industry and the epitome of its practice, obliges the commodity system to resolve the following contradiction: The technological developments that objectively tend to eliminate work must at the same time preserve labor as a commodity, because labor is the only creator of commodities. The only way to prevent automation (or any other less extreme method of increasing labor productivity) from reducing society’s total necessary labor time is to create new jobs. To this end the reserve army of the unemployed is enlisted into the tertiary or “service” sector, reinforcing the troops responsible for distributing and glorifying the latest commodities; and in this it is serving a real need, in the sense that increasingly extensive campaigns are necessary to convince people to buy increasingly unnecessary commodities."
- Guy Debord, 1967
Robots are not taking over your world -- just some jobs which mean the world to some people who have those jobs!
Even according to the timeline in the article itself the Swedish Electrolux Trilobite [1] was already press released when iRobot had their first proof of concept.
It's going to be the same with robots. Maybe actual lathes will disappear, but there will be computerized things that will perform that job, and people manning the computers. Some of the jobs will be automated (think Excel macros), but many will require human interaction and iterations.
"The more a robot looks like a robot the less useful it is."
Take fast food as an example: what's really going to change things are automated systems that reduce the number of people needed to run the place. Look at things like touchscreen ordering systems that cut the number of people talking to you. That's not a "robot" but it is automation of a sort.
At my supermarket, they replaced 2 regular checkouts with 8 self-checkouts in the same floor space. During rush hour, this greatly increases throughput despite actual scanning time perhaps being slower.
At McDonalds, it means they can higher fewer staff which means a cost efficiency.
Robots taking over the world may not be eminent. However, greedy people taking over the world with the help of robots and AI is already being facilitated for quite some time. Governments acquiring technology to hush the oppositions and free speech around the world is a common phenomena around the world. Usage of data along with AI to pinpoint, target and destroy enemies is being used for over a decade.
It is the usage of data and AI by the greedy community we need to be more concerned about, not silly imitation of life that barely perceives the world.
I am worried about something like this, which, while a parody, is not long off:
https://www.popularmechanics.com/technology/robots/a29610393...
Tl;dw, mini weaponized AI drones get loose and mass havoc ensues. I'm sure the feasibility behind these isn't close but still, it's an interesting design fiction.
Don't get me wrong; its terribly useful. I own a knockoff. But not a model of what robots could be, today. Think of facial-recognition drones delivering voice messages to people in a crowd. Think auto-drive cars and the terabytes of data and kilowatts of energy they command. Think of the Netflix algorithm for suggesting movies. Lightyears beyond Roomba.
Yes, automation will eat our lunch. Because automation is eating our lunch. Factories, city planning, scheduling, diagnosing - all sorts of automated processes are in industry and consumer space right now. Just ask Siri about it if you still have doubts.
As for the model/map, how does it sense location? Dead reckoning? Not a very significant feature if it can't find itself on the map, nor use the map to choose a route. Which I doubt (from videos of the Roomba in action)
Disclosure: I did an internship at iRobot a few years ago, although the overwhelming majority of my work wasn't related to Roomba products.
[1] https://spectrum.ieee.org/automaton/robotics/home-robots/iro... [2] https://spectrum.ieee.org/automaton/robotics/home-robots/new...
Implying that is just "a big name" that means "something far less interesting than it sounds" is an uninformed opinion. Plus, even [0], which is 15 years old directly refutes your comments about the effect of lighting. Lighting is helpful because it illuminates the scene that a camera is seeing, in the same way that it's easier to go on a night hike on during a full moon. We don't reference our position off the moon, it just makes it easier to see where we're walking.
[0] https://ieeexplore.ieee.org/document/1570091
[1] https://patents.google.com/patent/US9286810B2/en
According to the materials released by Qualcomm in 2018 (by which point, the Roomba i7 would have already have had lines in that plant), the processor used is a ARQ8009, a quad-core ARM processor commonly used in smartphones[0]. It's tough to find prices because the contracts are negotiated individually, but I'd put money on those chips costing much more than $1.00/piece.
Comparing VSLAM to lidar in any way besides saying that they technically can both be used to measure movement is also incorrect, and to actually call it a "kind of a cut-rate lidar" can only be due to a fundamental misunderstanding of VSLAM, lidar, or both.
Lidar measures distance along a single axis, with no additional information provided. You get movement by measuring changes in position by directly updating the distance value and inferring position change. It can be made robust by adding multiple lidar sensors, but is subject to weird edge cases with reflective materials or sunlight.
VSLAM only can measure change in vehicle position through the estimated position changes of significant image features through time, and provides no information about vehicle distance to the surroundings while the robot is standing still. You'd require stereo vision for that, and that does require more processing power, typically including a GPU like the one included in the NVIDIA Jetsons. You'd also typically use sensor fusion with an IMU, odometry, etc. to help with that.
Without stereo vision, you need motion to estimate position, which fundamentally makes VSLAM different from lidar, and makes your claim that it is "kind of a cut-rate lidar" completely false.
I'm not sure which DSP acronym you're referring to as a budget version of image processing , but if it has to do with digital signal processing, then yeah, of course. That's kind of the definition.
[0] https://www.qualcomm.com/news/onq/2018/12/17/help-qualcomm-t...
Clearly, VSLAM is being touted as 'vision' when in fact its far less than that, less than LIDAR even, which was my point. Which I feel was completely understood, yet still a silly attempt to shame me was spewed out saying "But its less than LIDAR" which was true, what I wrote, and beside the point.
VSLAM is computer vision. Full stop. Here's an article about it in a Springer Computer Vision guide [0]. I don't know what your personal definition requires for that to be the case, but if it doesn't include point recognition and persistent tracking by an integrated RGB camera, then I think it's unreasonably out of line with that of the rest of the field.
Which, just so we're clear, is the eighth reputable primary source I've presented you with directly refuting claims you've made.
Yes, I understood your point that you feel VSLAM is less than lidar. My point is that it's like stating that the moon is purple. Comprehensible, but wrong. And not in a "beside the point" way, but in a way that directly laid out why a core piece of your argument regarding how you try to compare VSLAM with lidar is invalid. Pretty sure that's not pedantry.
[0] https://link.springer.com/referenceworkentry/10.1007%2F978-0...
In the scheme of computer vision sophistication, its right down there with a fly's eye. That was my point, and I believe it was well understood. The pedantry of insisting it is computer vision full stop. Well, yes, about 1% of what computer vision could be. Niggling over that point is what's annoying. Instead of a good-faith discussion of how primitive this is, how tiny a feature it is, how miniscule the benefit it brings the consumer. How its pasted on the Roomba feature list because, hey, cameras are cheap and we can sell that feature for maybe $100 because the consumer doesn't know.
Also, I think you're underestimating both how valuable and complicated VSLAM is. Sure, it's not a multi-billion dollar state-of-the-art composite deep learning model trained on thousands of hours of data running on custom silicon, but it's hardly "This area is dark, and there's a bright thing over there!" Even if VSLAM only included visual odometry, that would still be a major improvement over other options in autonomous navigation, which is absolutely key in other domains that you seem to consider state-of-the-art. It's the same technology, and investments therein that make things like the RangerBot possible[1]. In a competitive market where BOM cost is a serious consideration, if you could get better results by not using VSLAM and relying only on something like laser odometry instead, that's what would be done.
Also, I'm honestly a little annoyed at your implication that I'm the one not having a good-faith discussion, while you have blatantly and repeatedly made false statements that could have been figured out by doing your own research, notable by your failure to provide any actual sources, instead making tangential anecdotal references to try to back up your statements. Heck, let's not forget your initial claim wasn't "limited vision" (which I debate anyway), but "no vision", and "no model of their environment" which again isn't simply true.
[0] https://store.irobot.com/default/roomba-vacuuming-robot-vacu...
I haven't touched this area since 2016, but I remember coding swarms of of robots that could manage to do very complex things from simple rules. In addition, they would be quite robust as they made very few assumptions about their environment / problem space.
You can see it from his list of failed companies, where only Unbounded robotics(smart robot hand), Rethink(collaborative robotics), Blue workforce(collaborative), Aria insights(drones) have seen their work continue successfully in their industry.
But yes, robots for the home are much harder.
So maybe robots won't take over the world, but they might take over businesses.
Perhaps this is because the business is already slave to its function. The exact separation and makeup of businesses reflects efficient ways to split up problems and integrate with other businesses. It's possible to refactor that given a robot that can perform a certain kind of action that is not directly equivalent to one person's job right now.
But my home is my home, it serves me, I do not serve it. For instance, one could imagine redesigning clothing to make robotic cleaning and folding of clothing feasible. I'm not going to do that! But redesigning PPE for that process in a hospital is not at all unreasonable.
Anyway, it's an apples to oranges comparison. Boston Dynamics are building incredibly expensive and sophisticated AI systems that will be first of interest to the military, who have very deep pockets and can afford to spend millions if it is deemed a good use of the money.
Roomba are building stuff that can be bought by the average person on the street.
I think the article provides an illustrative example of the different factors and complications that make it so difficult to go from a research prototype to an actual consumer product (a process which seems to take about 20 years on average). As usual, the lackadaisical improvement of battery technology is one of the central bottlenecks.
Of course, if your main interest is industrial robotics, where DRL approaches are more likely to become mainstream in the near-term, insights related to consumer products might not be as relevant. But note that industrial improvements, while important, will be mostly invisible to consumers except in terms of prices, unemployment and increased customization options for products.
Even then, the article can provide some perspective. Our processes (whether assembly lines or household vacuuming) are already highly optimized for a particular way of doing things, and researchers in particular tend to underestimate the amount of effort it takes to change systems to accommodate a slightly different method. So any new system not only has to be competitive with the existing one in just about every parameter, it has to justify the often huge costs of small alterations to the process. Thus, even in industry, getting to mainstream adoption of RL will likely take quite a while.