Fake it until you brake it
#21 | On why we should take robotics seriously
Robotics is a delicate topic for me, because I’ve spent so many sleepless nights on it. That’s why I don’t like to see so much attention on something I believe doesn’t make sense, like humanoid robots.
My amigo Chris edited this version of Unbeaten Path while trying to decipher tariffs with The David McWilliams Podcast. It’s less to do with economics and more to do with culture wars.
Finally, let me say this: No CEO with an ounce of intelligence will invest capital to buy thousands of humanoid robots to work in their factories or warehouses. They will redesign the process instead to eliminate the use of humans altogether.
Brittain Ladd is very opinionated with the added bonus that’s he’s often right. The comment above is what many I know in the robotics industry think but maybe lack the bravery to share with such a big audience.
Why? Well, what if they’re wrong? Being wrong is bad and doing so in public is horrible. Sadly for venture investors their opinions are expressed via their investments, no matter how discreet they are. There’s no way out of it, so this is what I think of humanoid robots.
They don’t make much sense.
But then, why is the industry - or at least the loudest voices - focusing so much on humanoids?
Jordan Nel nailed the reason in his piece “Robotics: A Product Selection Problem”: the TAM for humanoids is huge, literally every human activity.
Today they are starting with boring warehouse automation tasks because they can operate in a safe environment, without humans, on tasks that are not safety-critical and can provide valuable feedback to engineers. The real goal, however, is to bring these robots in the real world, replacing waiters, nannies and even postal workers.
I don’t buy this vision of the future.
Those who subscribe to this vision are committing the same error as the founders of Humane. They believed that what their small bubble of rich, tech enthusiast, Silicon Valley-based, and Twitter-educated people thought applied to broader humanity. We should never forget that iRobot’s Roomba, a very basic robot doing only one thing, that can be bought online for a few hundred bucks, that can save us from the hated task of cleaning the floor, sold “only” 50 million units since 2002 and the company was valued at $1B during Amazon’s failed acquisition. This isn’t bad in absolute terms, but it’s tiny for a consumer product, a leader in its space and sole player for a long time. For comparison, Apple sold 2.3 billion iPhones since its launch in 2007.
The belief in such an immense TAM is however very appealing for a certain type of fund, those that have billions of dollars to invest and need such humongous outcomes to justify their existence.
The reason why I am bothered is because I think we are missing a big opportunity here.
We are in a special moment in history. A period of change, a time of disruption. Industrial tech is becoming strategic and robotics can play an important role in making the West competitive again with China (as shared here, and here, and here).
Increased intelligence is one of the inflection points that will allow startups to disrupt large markets dominated by slow incumbents.
If we dedicate too much attention and money to the wrong applications of robotics, we could waste the momentum and miss the opportunity to use them for what they do best: dirty; dull; and dangerous jobs.
Ladd’s earlier comment is from a post where he opined about BYD’s new gigafactory, which is supposedly bigger than the whole of San Francisco.
BYD and many other Chinese manufacturers build such huge factories because there is value in vertical integration. These factories are highly automated and employ many robots. In some sense, they all learned from Tesla. When watching the infamous fly-through of Tesla’s gigafactory the thing that strikes me the most is how little welding there is. One of the most valuable innovations that Tesla brought to the car manufacturing world is the Giga Press, a gigantic, single-purpose robot that shapes metal sheets in one go. After Tesla introduced this technology in their factories, Chinese manufacturers quickly did the same and now also traditional carmakers are adopting the tech.
Do you see the parallels here?
For decades, we have produced little pieces in separate processes (and often in different parts of the world), welding and assembling them together in a long and inefficient process. We could’ve replaced humans with humanoids in those same production lines, but that would have been completely stupid as we would have been optimizing for a local minima - i.e. the human-centered factory - instead of escaping that local minima redesigning the whole process in a better way.
The key to making robotics work is starting from the problem, not the process. Processes are just solutions to problems. Processes can be changed if there is a better solution to the problem.
A typical (and overused) example here is the car vs horse: the problem is to move people faster and more comfortably, and the solution is changing the mean of transportation and not making horses faster.
To generalize, there are many problems that can be efficiently solved by specialized, product-centric robotic solutions. If a solution is offered at the right price, it solves a problem that’s big enough. If it actually works, customers are very happy to change their processes to adopt it.
Solutions like Autostore, Exotec or Kiva, for example, changed the way warehousing looked like. They did it by thinking about fundamental problems first, and then about the technical solution to optimize that process.
In other words, why should I use the same humanoid robot to assemble a car’s engine, prepare a poke bowl and sort goods in a warehouse? When all you have is a hammer, everything looks like a nail. If you massage the topic enough you can find similarities between the above-mentioned processes. However, the real world is different, these processes are very different and the right approach, in my opinion, is to have vertical solutions building processes tailored to the problem at hand.
Isn’t this the way it has worked for a long time? Yes, it is - but there is a caveat. Today intelligence is cheaper, more flexible, and easier to deploy.
I started working on autonomous driving in 2017, and I vividly remember when Wayve made waves (no pun intended) in the industry by claiming that they could do autonomous driving with end-to-end AI. People thought it was impossible, that it didn’t make sense, that “those who knew how to do it” were doing it in another way. History tells us Wayve was right: now everybody - Waymo and Tesla in primis - adopted the same approach.
Why did Wayve’s approach win? Because it was cheaper, more flexible and easier to deploy. All of a sudden, you can train your car in London and drive the next day in Berlin and Los Angeles.
The same thing, I believe, is happening to other areas of robotics. Today, robots are able to understand the real world way better than they could do before, and all of a sudden it’s not unreasonable to think about robotics deployed by traditional SME, increasing their margins and optimizing their processes. The vertical robotics solution - purpose-built with specific problems in mind - can now give superpowers to traditional industries, which are those most affected by aging population and lack of people willing to do the job.
One word of caution is needed though.
After Nvidia GTC, the VC world has been flooded by the concept of “Physical AI” and many software investors that missed the “generative AI” train boarded this one as if it was the second coming of Christ. The problem is that the real world is not software, and “good enough” in the real world doesn’t have the same meaning as good enough in the software world.
At my first startup, we were building autonomous robots and we had a robot that was running around our office autonomously every day, the whole day. Once in a while something minor would happen, but we were happy anyway: it was good enough for us.
Somehow we managed to convince a very very big US retailer to test our robots, and soon we understood what “good enough” means in the real world: >99.9% uptime. Needless to say, that wasn’t even close to what we had, I wouldn’t be here writing this post otherwise.
This little tale taught me something: in robotics, your competition is usually physical labor, which generally works well enough. The question “Does it work?”, in robotics, has a whole other meaning than in software.
“Does it work?” is the first question one should ask when doing due diligence on these companies. It shouldn’t be answered superficially.
If you’re building a full stack company, your TAM, value proposition, and customer “orders” will typically be obvious. But, we care about performance. That’s the hard part. And that’s why the opportunity for full stack remains massive — and underpenetrated.
Investing in the real world is a huge opportunity, but building in it is very delicate.
As our industry is finally getting excited about the built world again, we shouldn’t waste this opportunity - it’s too important that we modernize our industrial base, and robotics must play a role in it.