Alphabet Shakes Up Its Robotics Division | Hacker News Reader

Alphabet Shakes Up Its Robotics Division

nytimes.com

73 pointscryptoz70 comments

Perhaps we are just impatient.

Karl Benz invented the 'first automobile' in 1885 [0]. 29 years later, in 1914, Ford built the first mass-production factory for his autos. It initially produced 4 cars/hour, an annual rate of about 15,000. In 1915, there were 8,000 cars in all of the US. [1]

I cannot find data, but let's guess that it's 10 more years before 1% of households in the US have cars. That would have meant about 500,000 cars.

1885 - 1925 is 40 years between invention of the first gas powered auto and 1% market penetration (bearing my assumption). Where was the robotics industry in 1976, 40 years ago?

Moreover, the automobile solved one problem: Moving a human from point A to point B on a paved road. Many efforts are approaching robots as consumer products by trying to solve many very hard problems all at the same time. Perhaps robotic vacuums are the (only?) successful commercial robot because they solve a similarly discrete problem.

It is understandable that we have not yet seen the type of commercial breakthroughs that we so highly anticipate. Give it another 20 years...

[0] https://en.wikipedia.org/wiki/Car#History

[1] http://www.thejournal-news.net/opinion/commentary/here-s-loo...

The pace of technology adoption has sped up but that's comparing electronics to, in many cases, large appliances so I'm not sure how relevant the comparison is.

https://hbr.org/2013/11/the-pace-of-technology-adoption-is-s...

is an interesting graph. Historically a lot of consumer tech took a long time to be adopted.

A nice segue is that especially in manufacturing robots are plentiful. From pick-and-place units that move too fast to follow with the naked eye to welding robots and so on. In a way a current car manufacturer is a series of giant robots glued together with a relatively small number of humans.

Universal robots, ones that can do basically any task that any human can do are not even on the drawing board right now, so we're limited to these application specific models but where they are engaged they are incredibly effective and have all but replaced the humans that did those jobs previously.

The reason often quoted is that this is because robots can work 24x7, won't unionize and so on, but the little fact that people usually miss is that repeat accuracy is a very important factor in reducing part and product failure so betting on robots means that you get a much more reliable and consistent product out the door saving you a fortune in re-work and warranty claims.

"the little fact that people usually miss is that repeat accuracy is a very important factor in reducing part and product failure so betting on robots means that you get a much more reliable and consistent product out the door saving you a fortune in re-work and warranty claims"

That is a very good point, the other factor is that robots can collect massive amounts of data faster and more accurately than their human counter parts.

One data point:

"1974: The world's first microcomputer controlled electric industrial robot, IRB 6 from ASEA, was delivered to a small mechanical engineering company in southern Sweden"

Here's a cute video showing this robot, still in use 40 years later:

https://www.youtube.com/watch?v=2xNgQhLAPyI

The other thing with vacuums is that for many people, especially those with kids and pets, vacuuming and sweeping is a pretty frequent task that usually takes more than a few minutes.

Mowing lawns is the other obvious task that falls into this category and, indeed, robotic lawnmowers do seem to be coming onto the market.

>I cannot find data, but let's guess that it's 10 more years before 1% of households in the US have cars. That would have meant about 500,000 cars.

Ford produced 500,000 Model T's from 1908-1914, and 500,000 in 1916 alone. While they didn't all sell in the US, with 1.3 million T's produced, it's safe to say they sold half a million cars in the US by 1916.

That bumps your 40 year estimate down to 30.

I have to say I'm kind of baffled about all these companies complaining "Regulation is hurting self driving cars!" and "Regulation is hurting drone delivery!" when there are pretty much zero barriers to enter the consumer robotics market for a machine that dusts your house, does your laundry, does your dishes, picks up clutter from the floor, etc.

(And yes I realize these tasks involve surprisingly hard AI problems, but why isn't Google already running pilots where they give people special washing machines, laundry machines, robot-friendly book shelves, etc etc that help mitigate the hard AI aspects of the problem?)

Simple: they lack leadership. A proper leader would focus on setting the framework for autonomous machines to operate. Without leadership, all they can do is showcase shiny greenfield projects.

The same reason they haven't given everybody "pilot" autonomous cars. Expectations are huge now, everything must be perfect.

I agree. What about a robot that simply move clothes from the washer to the dryer? That's a pretty well defined tasks and is basically 50% of the work of doing laundry.

Someone else fold your close for you? Moving the clothes from the washer to dryer takes 15 seconds.

You have to be there to do it, which interrupts whatever you were doing at an annoyingly pecise time.

Clothes can sit in the washer for hours. :-) I transfer when I'm walking by to go to the bathroom. Actually, with my AppleWatch, I usually unload dishwasher, washer/dryer when I'm told to stand up and move for a minute.

I actually have a superpower where I can smell mildew when no one else can. So even 30 minutes of leaving the clothes wet is too much for me.

It's interesting when I meet people I can smell who has left their clothes in the washer too long.

I live out of the dryer :-(

All-in-one units exist that just have a single front-loading tub. It seems they're designed primarily for where space is tight, so I suspect there are some tradeoffs.

But as someone else said, that's probably the step associated with doing laundry that's the least effort. Putting clothes away is the biggest effort and good luck designing a robot to handle that.

In 'serious academic' robotics, if you only have one hokuyo urg lidar [1] and one xsens mems imu [2] you're on a low budget project and you'll probably need a bunch more sensors before your product can operate reliably unattended. But you've already spent ten times the cost of an entry level roomba.

The kind of butler robot we can make today [3] would seem extremely expensive for the functionality it could offer, making it hard get the sales needed to recoup your up front engineering costs - let alone to achieve the economies of scale needed to bring down the price on those sensors.

[1] http://www.robotshop.com/en/hokuyo-urg-04lx-ug01-scanning-la... [2] http://www.newark.com/xsens/mti-20-2a5g4/mems-module-3d-moti... [3] https://www.youtube.com/watch?v=tODGHw8uNnY

I misread that as 'butter robot'.

https://youtu.be/8d_hveJL4mU

Thanks for the links; I love this stuff. I can see the laser rangefinder [1] being expensive. But what makes the xsens gyro/accelerometer [2] so expensive compared to a smartphone sensor, and why does a robot need that?

Yeah that thing seems way overpriced, or at least far too high-spec for most applications. Quadcopter flight control boards containing gyros and accelerometers can be had for less than $20, and those do just fine.

Accurate heading info requires good gyros and is moderately expensive.

It's a lot cheaper than it used to be. When we did our DARPA Grand Challenge vehicle ten years ago, we had about 3 degrees of heading noise, which kept the system which integrated LIDAR data into a terrain map from reliably mapping. The system had a MEMS gyro with too high a drift rate, and a magnetometer compass for long-term correction. But there were enough magnetic fields from the onboard equipment to mess up the magnetic compass. The software could detect trouble, and would stop, tilt the laser head to scan the terrain ahead and get a clean image, then advance. So the vehicle kept stopping to rescan.

We looked into better heading gyros, but they cost about $20K at the time.

If you think that's expensive, have a look at an OXTS RT3000 [1] where you can spend closer to $20,000 - that's the sort of thing you buy for your prototype self-driving car. Or at least, it was when people were competing in the DARPA Grand Challenge eleven years ago.

Applications for IMUs include dead reckoning (working out your position based on rotation and acceleration measurements), orientation sensing (figuring out which way is down is much easier than dead reckoning, because you can just sense gravity), and compensating sensor readings for robot motion.

Some cheap sensors aren't very precise; you won't notice if your phone thinks it's rotating two degrees a minute around the vertical axis, but a robot will if it moves for five minutes then it's ten degrees off course. I haven't looked at the market for a long time - I suspect the cheaper IMUs are better nowerdays. Eight years ago they didn't offer very good precision and the open source software could be very fiddly to get working, if you could get it working at all.

[1] http://www.oxts.com/products/rt3000-family/

At the 2004 DARPA Grand Challenge, the IMU/GPS unit that CMU's team used required 4U of rack space and air conditioning. There's been a lot of progress since then. The ADXRS642 MEMS gyro[1], which costs $50 in quantity, has a drift rate of 20°/hour. Ten years ago, you were lucky to get 20°/minute from a MEMS gyro. If you're doing sensor integration while moving, you need a stable orientation reference. Alternatively, you can use SLAM to process video and align to the world, which works well today.

The OXTS RT3000 has impressive specs - 1cm accuracy in world coordinates. That's better than you need for most robotics applications that aren't running blind. The Trimble AP10 [2] has the same functionality at much lower cost, but is not as accurate.

A lot of progress in robotics comes from component improvement like this. There was a time when just getting a wireless data link to your robot was a big deal.

[1] http://www.analog.com/en/products/mems/mems-gyroscopes/adxrs... [2] http://www.trimble.com/gnss-inertial/AP10.aspx?dtID

If you need expensive sensors and other parts, the unit economics for domestic robots don't make sense; a single robot would cost too much for any single family to justify, compared to just hiring someone to do the laundry once or twice a week. A delivery drone, on the other hand, would be used all day by some business, so a large upfront investment makes sense.

I am reading this while my neato robot is disoriented trying to find its home base and is bumping into every wall on the way.

I basically have zero confidence that we'll see something like what you are describing in next 2 decades atleast.

Really regulation is critical for the success of self-driving cars. Nobody is going to want to buy self-driving cars if there are ones that get into lots of crashes so I think we're going to get into the same situation we have with the FAA where the big players all want the regulator to have real teeth.

Consumer robotics doesn't have much of an entry barrier but as other have pointed out the components you need mean that prices are going to be in the $X0,000 range. That's too much for consumers but I think we're going to be seeing robots in warehouses soon.

"Really regulation is critical for the success of self-driving cars. Nobody is going to want to buy self-driving cars if there are ones that get into lots of crashes..."

If I were a car manufacturer and nobody wanted to buy my cars because they were seen as dangerous, I might try to make my cars safer to increase my sales, even if the law didn't require it.

The problem is trust. The car being safe and being marketed as safe are different.

You might, but that costs money, and other market players are always invented to undercut you.

Because the line of how much is too much is indistinct, someone is going to guess wrong before too long. This is even before you mix in perverse organizational incentives involving short-term view or individual profit vs long-term company health.

Complaining about regulations holding you back is a great way to hide that your technology isn't actually there yet. It's marketing.

Companies always love to complain about too much regulation tho.

There's one very simple reason that general-purpose robots have failed to become economically viable: robots have to compete with humans, and humans are really cheap.

I also think that not enough thought has been put into the question of: if we do manage to replace unskilled workers with robots, what do we do with all the people who are only capable of unskilled work? Not everyone has what it takes to become a coder.

https://www.youtube.com/watch?v=7Pq-S557XQU

That's basically the theme of The Second Machine Age. There are a lot of people around here who IMO are way too optimistic about the timelines for fully autonomous robocars. But if/when such a thing actually arrives, the effect on employment will be huge. Even if it may well still make sense to have a human present for, say, package delivery, the number of people whose job is primarily to be a driver is huge.

I don't understand what you mean by humans are really cheap, you can't upgrade a human , you can't make him faster, you can't install latest update or add some components, the main disadvantage of robots nowadays is that we have not yet achieved a human level AI that gives us the abilitiy to replace humans by robots, but if we ever do that or even achieve just good advancement in it, I strongly doubt that any "unskilled" human worker would be better or compete on the job.

Well, the cost of the capital outlay for reasearch and components, along with the ongoing maintenance costs for the robotic replacement should outweigh the cost of training, hiring and paying the human worker for the same job/task. That is what is meant by the "general purpose worker" as a human being cheaper than a robotic one. And is also why most robotic replacements, as far as I'm aware, have been very specialized and in extremely repetitive tasks that require minimal "visual cues".

> I don't understand what you mean by humans are really cheap

Seriously? I can rent a human for $7.50 an hour. In other countries they're even cheaper. If you want to sell me a robot to clean my house, that's your competition.

That was not a real statement, I just meant that it made contradiction in my mind, not speaking on short run, but in my country and european countries, you need to pay a lot of money to hire someone that includes health insurance something similar to US 401K beside so many things, so it seemed to me like on the long run, robots are much more cheap.

> "so it seemed to me like on the long run, robots are much more cheap."

Yes, in the very long run, assuming relatively unchanging tasks and capacity demand.

But most jobs aren't like that - industrial robotics are dominated by robots that are purpose-designed with a singular task in mind, with a narrow envelope for deviations. A human can be re-trained to do something completely different than before, whereas a robot would likely have to receive completely new hardware. A welding robot cannot become a precision fan-balancing robot, but a human can be retrained from one task to another.

It's cheaper to train a large group of humans to do something different than it is to upgrade the hardware & software on the same number of robots.

This means that some jobs (see: automotive manufacturing) lend themselves easily to robotics - the throughput is steady, the capacity requirements are stable (i.e., you don't suddenly need 5x more cars in July than in March), and even between different models the general process is the same.

Other jobs (see: warehousing, shipping) do not lend themselves to robotics easily - the job changes regularly as processes evolve, the capacity requirements are chaotic (i.e., you need to scale to 10x your normal throughput during Xmas), and there is significant variation in process requirements (e.g., Japanese retail warehouses that run vertically instead of horizontally).

Robots are also still astonishingly expensive and largely made-to-order with long lead times. This means that it takes many years to break even on a robot (whereas you can fire the human when you don't need them), and you can scale easily with humans (it's much quicker to find a human than it is to order and receive a robot).

I'm just going to repost what I wrote 17 days ago:

"Why are they against militarizing robots? It pays the bills. They could start with consumers or business but it's harder to build a market. This makes it harder to fund further development. Once the technology is developed by anyone for non-military use, transitioning to the military is easy. There's not some magic that prevents consumer developed technology from reaching the battlefield."

It pays, but with heavy constraints. You are subject to cycles in the military budget and to the government sales process.

I highly doubt Google is interested in that revenue stream. Google doesn't just want to pay the bills -- and it doesn't need to. It wants to develop entirely new businesses that complement its existing ones, e.g. Nest.

Also, it's bad for the brand. There are probably not many companies that sell both weapons and consumer goods...

I believe the idea in the new model is that each unit does need to pay the bills.

It doesn't have to be that way. Alphabet can pump money into the robotic subsidiary forever. But it'll be more and more obvious how much of a drain it is when the robotics company is a black hole of profits.

Weapons and business goods at scale may be more viable - a la Boeing.

> Why are they against militarizing robots?

I get that their duty is to their shareholders, but you have to draw a line at some point. I wouldn't want to work on tools of violence.

It seems like every consumer and industrial robotics company, with the exception of the drone makers, tends to run into trouble sooner rather than later - and it turns out being a division of Google doesn't help that much.

Rethink Robotics for example had a great start, but soon ran into cultural issues and weaker than expected sales. iRobot has been unable to escape the vacuum cleaner market. KUKA and the other really big industrial robot manufactures have started doing new things with sensor-enabled manipulators, but the customers haven't exactly flocked.

What is it about robotics that seems to run these companies off the rails? I worry about Fetch robotics hitting a similar barrier, despite having a pretty clear business case.

> iRobot has been unable to escape the vacuum cleaner market.

In the consumer space.

iRobot has a healthy relationship with the Defense Department.

http://www.irobot.com/For-Defense-and-Security.aspx

That's true. In fact, it seems like cosying up to DoD is the only way for a robotics company to really get anywhere or stay afloat in the long run. This shouldn't be true, yet seems to be. Why?

Robots are expensive and the DoD has nearly infinite funds?

While it's true that DoD can allocate considerable funding, it seems strange that there is no other business model.

The DoD abhors intelligence in robots. They need a machine that's just smart enough to do a specific job -- and no smarter. They're willing to pay for the smarts -- and may accept compromise for a solution that gets them most of the way there. That makes filling their orders easy compared to the consumer market, which wants mechanical butlers for cheap.

From my (probably naive) point of view, robotics sits on top of three pillars, mechanical engineering, electrical engineering, and computer science.

To even get off the ground, you need strong people in each of the three pillars. Additionally, you need a leader who is sufficiently knowledgeable across all three pillars. Purely from an HR/talent acquisition perspective, this is difficult logistically and financially. Uber raided NREC to accomplish this, essentially creating a second headquarters, and paid quite a bit to get the talent.

Then, you need a prototyping facility for both the ME/EEs.

To start from scratch, a very hefty investment has to be made before product development has even begun. Then you're bleeding money through the product development phase (hopefully with some pre-sales, but odds are pretty low of this). After product development, you hit manufacturing, which is extremely difficult. Bolt.io has some good reading on this.

To get to a product driving around, you've spent millions over the course of a year or two. To amortize that cost in a reasonable fashion, you've got to produce thousands of units. Fanuc took over manufacturing early on and was able to accomplish this, but for a lot of companies, the sales volume just isn't going to be there.

The DoD, with SBIRs and other long term contracts, helps fund a lot of the product development, often provides you with a paying customer, etc. The risk and amount of monetary bleeding drops significantly as soon as you partner with the DoD.

I absolutely agree about CS and MechE but I my experience is that you can generally do just fine without anything cutting edge in terms of electrical engineering until you've grown a fair bit. Depending on what you're doing you might not need anything but off the shelf components like the hands we sell[1], the arms Universal Robotics sells, etc. Or if you're doing something more advanced it's probably using standard motors and ICs with a custom PCB. It's only when you're getting more advanced that you want more control of, e.g., the power system for production quality robots.

[1] Shameless plug http://www.righthandrobotics.com/

> That's true. In fact, it seems like cosying up to DoD is the only way for a robotics company to really get anywhere or stay afloat in the long run. This shouldn't be true, yet seems to be. Why?

I don't see the problem with it. It's not like YCombinator or other VCs are going to pay for years and years of research.

The DoD can afford to throw money at robotics projects and fund the grunt work of getting the technology developed. Spending a billion dollars researching and developing a walking robot or whatever is a great use of money if it means keeping a bunch of humans out of dangerous situations.

For everybody else it's a cool idea, but not useful enough that they'd pay for it. Better to let DoD pay for it, and wait until that technology makes it into other products.

The vast majority of robots are solutions looking for problems. The problems in the DoD are both pressing and obvious, making this problem show up less often.

Savioke and Fetch have it right. Solve small problems before moving on to bigger ones. The other robotics companies are simply realizing that now. Alphabet doesn't need to, they are attempting to reach perfection in a void. All the power to them!

I suspect that in the consumer space specifically there aren't a lot of obvious tasks that are amenable to being done by robots. Vacuuming is one of the obvious ones--which is why we have Roomba etc.--but even that requires fairly specific types of layouts. Laundry, cooking, loading and putting away dishes, general cleanup. We have appliances to help with some of that and we can buy, say, food that's prepped to a greater or lesser degree. However, pretty much none of those are things that one can imagine a practical consumer robot doing.

I would buy a robot that could fold clothes, but that turns out to be a really hard problem (http://www.wsj.com/articles/why-robots-still-cant-fold-your-..., although there seems to be some progress in the lab: http://www.telegraph.co.uk/news/worldnews/asia/japan/1192259...)

Well, and for most people, it's not even really a "folding clothes problem." It's take everything upstairs. Put these socks in that drawer. Hang this shirt on that closet. Hang or fold these pants and put them in the right place. Repeat for each item. It's actually a move a lot of things around the house and put them away in the right way in the right place problem. Oh, and do it cheaply, because very few people are going to pay thousands of dollars for your put clothes away robot.

There is Kiva systems. Their non-Amazon customers are sad but I think it would be fair to call the company a big success. I think that the warehouse logistics market is going to go big over the next 5 years. But of course I'm in a startup targeting that space so I better think that.

Kiva really seems to be the anomaly - and afterward, Raffaello D'Andrea went off to be a professor (though he has some new companies in the works to do with drones).

But their business model was proven and others have followed suit with arguably more impressive technology (http://clearpathrobotics.com/)

If you take a US perspective, perhaps, but globally, there are other bright spots. For example, after reading about iRobot's struggles to get permission for its wireless lawnmower robots, I was surprised to look into it and discover there were a fair number of lawnmower robot manufacturers for both large and small suburban scale robots which reportedly work well and are rather popular in the EU. Smaller lawns and higher labor costs speculated to be why.

(I liked the idea, but after calculating how big our property was, only the most expensive commercial lawnmower robot reached parity with our John Deere Z-track riding lawn mower's capabilities, and the payback period was implausibly long, so I'm stuck doing the mowing myself. For now.)

I believe a lot of the autonomous mowers also require you to lay a guidewire around the boundary of your property.

There are a few that don't, but they require your lawn boundary to be fully enclosed. They bump off the barriers to know to turn around.

I did look into this briefly, but even though I only keep a fairly modest portion of my property as a lawn, I'd still probably be looking at close to $3K, which would be about a 4-year payback relative to my lawn service--which also does some weed whacking and mows some small patches that aren't really connected to the rest of the lawn. Laying the guidewire would also be a pain.

But I can see the attraction if someone had a smallish lawn--especially if they didn't already have a lawn mower. It would be just one more task they wouldn't have to think about.

It is pretty magical having one - it is almost like the lawn never grows after it reaches a certain height. Also, the neighbors love asking questions about it.

There are also noise restrictions and lawn care differences (do you like stripes on your lawn?).

The two you mentioned (smaller lawns and higher labor costs) are the biggest drivers though.

I lead a team working on these for a major outdoor power equipment company, so I'm more than happy to chat.

The article doesn't say much. Google has two humanoid robotics companies, Boston Dynamics (near Boston) and Schaft (near Tokyo). Schaft has better actuators and Boston Dynamics has better control. There's no indication that Alphabet has been able to integrate the technologies. That was the whole point of acquiring ten robotics companies. But if they can't be brought together, the acquisitions were pointless. None of the companies have significant revenue or sales, so running them as a conglomerate doesn't do anything.

Alphabet is so secretive that no one can tell from the outside if they're doing something great or getting nowhere. Boston Dynamics stopped updating their web site after the Google acquisition. Schaft has no visible web presence at all. Raibert, Boston Dynamics' CEO, is 67, time to retire. Who replaces him?

What I want to know is if Alphabet is still trying to make robots for war. They promised only a year ago they wouldn't do that and wouldn't become a defense contractor, and then a month ago we found out they "failed" to sell their robots to the US Army.

Good thing it was only Google that had to abide by the "don't be evil" mantra (more or less). Alphabet apparently could care less about it.

Reply on news.ycombinator.com