Hacker News Reader

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.

It's not that the robotics is hard. It's hard to stop the time-traveling saboteurs from the Resistance from undermining you.

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).

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)

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.

I don't think it was ever meant to be live driving at highway speeds:

https://www.forbes.com/sites/stefanseltz-axmacher/2020/06/16...

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?

[0] https://medium.com/starsky-robotics-blog/the-end-of-starsky-...

>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.

More specifically, humans have an incredible high adaptability:cost ratio.

There aren't many businesses where precision:cost or volume:time are more important than labor costs.

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.

I've worked in software automation for about a decade now, and that's been my learned wisdom too.

Don't try and boil the ocean: see what COTS is available, adapt your process to be able to leverage that, plug it in, and move on to the next project

As commentor above noted, volumes have to approach obscene to justify a moderate+ amount of custom, one-off implementation work.

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.

> 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.

> 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/

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.

Saying the market isn't big is indeed questionable, as the total addressable market for advanced robotics is easily that of the global labour force.

Likely what is meant is the market for current state of the art robotics, which have limitations and are cost prohibitive (capital wise).

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.

[1] https://en.wikipedia.org/wiki/History_of_numerical_contr ol

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.

This is truly a crossing the chasm problem.

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.

Sadly, not that easy.

> 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.

You are running the joke by being that Obvious, Cap.

The fittest robot for any application is indistinguishable from an appliance or a machine tool made for that application.

If your robot can't receive either of those labels, your robot company is doomed to a slow death.

I cannot agree enough - we always used to say the best robots aren't called robots, they're called washing machines and dishwashers.

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.

https://www.youtube.com/watch?v=TUx-ljgB-5Q shows some footage of the robotics they use.

Tom Scott had a good video a couple of weeks ago about a grocery packing warehouse that has a sophisticated picking network:

https://www.youtube.com/watch?v=ssZ_8cqfBlE

That Ocado factory had a fire recently cause by a robot collision, so the quality doesn't seem quite there yet [0]

[0] https://www.euronews.com/next/2021/07/19/ocado-warehouse-fir...

And just today: https://www.aitrends.com/robotics/softbanks-humanoid-robot-p...

Looks rather small number of investment

There were hundreds of copycats startups in China following the trendy business ideas at the moments.

The Groupon era Streaming Short video Gif sharing Etc...

It just looks like not enough money in robotics, not that robotics are wasting them

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.

Reply on news.ycombinator.com