The First Tour of Giant AI’s Robot Lab

Visiting Giant AI

Visiting Giant AI is like getting a tour of a secret lab that shouldn’t exist run by an eccentric genius. The kind of which we remember from “Back to the Future.”

Adrian Kaehler is that genius. 

He built the computer vision system for the first autonomous vehicle that later became Waymo. He played a key role in the early development of Magic Leap, an augmented reality company that just won best of show at the industry’s biggest gathering, AWE (for Augmented World Expo). He also wrote what many say is the book on Computer Vision which is still used by many computer science departments. Today he is leading the Giant AI company which is building humanoid robots that can work on manufacturing lines, doing the same jobs humans used to do and many new ones. Giant is invested in by Bill Gates and Khosla Ventures. 

He saw the problems long ago that robots will bring. The earlier companies’ robots were designed and built to be very precise, which means they remain expensive today. You see many of these in factories today, they are heavy, don’t work well with humans, have to be programmed months in advance and are hard to retrain and don’t recover well when errors are made. Some are too dangerous to be around like the ones in Tesla’s factory in Fremont, which has some robots in cages to keep all humans away. 

He also saw the solution: AI that builds a new kind of operating system. One that learns faster than anything most humans could imagine. It learns so fast that you only need to show it a few times how to do something and it’ll do that thing from then on. One that enables new lower-cost components to be used. Ones that are less precise

When I watch the Universal Worker move, I can see how the tendons that make it work create a very different, animal, sort of motion. It is kind of springy. This would be a non-starter for a traditional robot, but the AI control, just like with a person, manages this and makes it all work out. Dr. Kaehler tells me that the use of this sort of tendon system is central to how the robot can be so light and dexterous, as well as why it can be so much less expensive than traditional robots.

 It’s the new AI that enables this new lower cost and safer approach. 

So, getting into his lab first meant a lot to me. Why? I think it means a lot to you, too. 

It means we will have to rethink work. From scratch.

Is your happiness and income coming from you pushing a button on a machine? Really? I worked on HP’s manufacturing line when I was a young man of 17. One of my first jobs was working the wave soldering machine there and shoving circuit boards into the wave, which instantly solderied the whole board. I had helped my parents and brothers hand build hundreds of Apple II. My mom taught us to solder. If you get good at it, like my mom was, you could do maybe a board in 30 minutes. I saw how manufacturing lines can change labor from my kitchen. My mom worked for Hildy Licht, who got hired by Apple to take on the task since they couldn’t make enough in its own factory. Apple cofounder Steve Wozniak, AKA “Woz,” told me that those boards had fewer failures than the ones made in its own factory. It also makes me Apple’s first child laborer (I was 13 at the time).

Anyway, I never wanted to do such a job again, given how boring it was. I loved that Wave Machine because it saved many hours of labor. I dreamed of a day when a robot would stuff the board too. I had to do that over and over and over again.

I wish I had a Universal Worker by Giant AI Corporation back then.

As he showed me around he was telling me what was making these robots so special. The AI inside is next level. See, I’ve been following AI since the very beginning.

I was the first to see Siri.

That was the first consumer AI app. I also have the first video, on YouTube, of the first Google Self Driving Car. Long before anyone else. That was the first AI on the road. I have been following AI innovators since the ve beginning.

This robot is using the next generation of all that.

Don’t worry, though.

You do get that we are in an exponential world, right? One where we need more workers, not fewer. Even if Giant got a huge funding deal, for, say, a billion valuation, it still couldn’t build enough robots to replace ANY human for MANY years. These are to fill in the gaps for when you can’t get enough workers to keep up with demand.

Anyway, back to the lab. Along each side I saw a row of robot prototypes for what Giant AI is calling “the Universal Worker.” Each was being tended to by staff, as Adrian gave me a tour he explained what each was doing. A new form of ML that uses neural radiance fields to see – the engineers are putting finishing touches on blog posts that will soon come going deep into the technology. In the video Kaehler goes into some depth about what it’s doing and how it works.

Each robot had a humanoid form. Even a smile on the face. And the head moved in a very unique way that I had never seen before. Strangely human like. Which, Adrian says in the video embedded here, is part of its ability to learn quickly and also get the trust of the human working aside it. It also lets it do a wider range of jobs than otherwise. It sees the machine, or task, it is standing in front of like we do – in 3D. And, in fact, there are many other similarities between what runs under robots, virtual beings, autonomous vehicles, augmented reality headsets and glasses. Kaehler is the only human that I know that has built three of those and he says that they all are strongly connected underneath in how they perceive the world, let others see the perceived and synthesized world.

As you get a look around his lab, you do see that they feel like early versions of the Tesla Autopilot system: a little rough and slow. Heck, even today, four years later, it does 6,000 more things, but it still seems a little rough and slow. The Universal Robots feel the same a bit to me. At least at first. Until I started watching that this was real AI learning how to grasp and drop things. It felt humanlike as it dropped a rod onto a machine yet another time in a row without dropping it. 

I remember talking to the manager of the Seagate Hard Drive factory in Wuxi, China, about why he hired so many women. Nearly his entire final assembly line was women, highly trained too, I watched several run a scanning electron microscope. I never will forget what he told me: “Men drop drive off line, pick it up, put it back on line. Women don’t do that. They bring over to me and admit fault.”

This robot was learning quickly how to recover from its mistakes. Which is how it was designed, Adrian told me. It has grids of sensors in each finger, which can do a new kind of “feeling” than I’d ever seen a robot use before. Each of those sensors was being pushed and pulled by a cord going to a machine in the belly of its humanoid form. On the end of an arm that was built from cheap consumer processes. The hand shakes, just slightly, especially if a big forklift goes by. 

Giant’s AI is what makes it possible to become far less expensive. It “sees” the world in a new way, using something the AI engineers call “Neural Radiance Fields.” A new form of 3D scenes that you can walk through. In Giant AI’s case it moves the hands through these radiance fields, which are unlike any 3D data structure we’ve ever seen before.

Its AI is constantly adopting and learning, which lets it figure out how to recover from a mistake very quickly. Adrian wrote the math formula on the board on a previous trip. It keeps pushing the hands toward the best possible outcome. So, you can slap them and they’ll recover. Or, if an earthquake hits and it drops your motor before it goes into the box it was supposed to put it in and the machine shakes. It still should be able to complete the task, just like a human would, or try to save the part, if possible, and if possible it will report a problem. 

Anyway, at this point, you are wondering “why did you hype up Tesla’s robot so much?” Last week I did. Those who are inside the Tesla factory tell me that their simulator gives them an unfair advantage and will let them build a humanoid robot that can walk around and do a variety of tasks much faster than people are expecting. You’ll see Tesla’s robot in September as part of its AI day announcements. Yes, hardware is hard, even if you have the best simulators, it is getting easier.

In a way this is a David vs. Goliath kind of situation. So Giant had to focus on a very specific, but large enough, problem: one of low-skilled workers and what they need help with.

Which is why Giant’s Universal Robot doesn’t have legs. It isn’t a trillion dollar company. It can’t afford to put legs on a robot that doesn’t need them. A worker in a factory always stays in the same place and does the same job over and over and over.

It doesn’t spy on you the way that the Tesla robot will (Giant’s AI only can “look at” the work surface in front of it). It can’t walk around your factory floor mapping it out, or watching workers in other parts of your plant as it walks around.

It also doesn’t have a completed mouth, or a voice response system, or the ability to really communicate well with other human beings the way the Tesla robot will need to do. Which makes the Giant robot far cheaper than the Tesla ones and it ready now, at a speed slower than human, or soon, at same speed.

That said, Kaehler is keeping up to date on the latest computer vision research and he knows that Tesla’s will do many things Giant’s can’t, and that’s fine with him. He doesn’t have a car company to gather data about humans in the real world. It isn’t his goal to build a robot that can deliver pizza. Just do boring jobs that humans need an extra set of hands to help do.

Giant AI already has orders, Adrian says, but the company needs funding to get to the place where it can manufacture the robots themselves.

I remember visiting “Mr. China” Liam Casey. I visited him in his Shenzhen home and he gave me a once in six thousand lifetimes tour of Shenzhen that I treasure to this day. Then he took me on an even wilder one over his homeland of Ireland, where he took me to a research lab that Mark Zuckerberg ended up buying.

What did Casey teach me? He had the same problem. No one would invest in his business, even though he had customers. How did he get his orders done, I asked him “I got them built.”

“But how? Did you have something to trade? A house, an expensive car, secret photos, what?”

“My US Passport.”

The factory owner demanded his passport in trade for building his order. A form of collateral I’d never heard of before. Then had Casey travel the country to all his factories to do a certification on each. That led Casey to see the power of databases, particularly ones for tracking supply chains. Which is why he is Mr. China today, and makes many products in his PCH company that probably are in your home today. He used that early research about China’s factories to become the supply chain leader that many technology companies use to build their products.

Giant needs the same today. A way to get the product finished and manufactured. Capital, and lots of it, to get to where these are working hard to make everyone’s lives better.

Tesla’s simulator has ingested a lot more than just where has the car gone. It knows EVERYTHING about its owners. So, when an engineer wants to recreate a street, it is amazingly real and the people will even stop to say hello or let you check out their dog. Then you can make it rain. Or make it sunny. Over and over and over again. 

Why is it so magical? BECAUSE OF the data the car and phone collects. A Tesla crosses the Golden Gate Bridge every 10 seconds. No one else is even in the same universe in data collection capabilities.

Tesla has a similar bleeding edge AI to Giant’s but Tesla’s has billions of times more data than Giant ever will get its hands on.

However, do you just need a machine to push a button or two every minute or two it notices a job is done or do you need Tesla’s AI and simulator that will have to do a whole lot more? No, at least not now, because the costs will be completely higher for the Tesla robot, which will need to walk and get in and out of autonomous vehicles.

That said, now that I’ve seen the Giant AI and how sophisticated it is with literally no data when compared to the Tesla system I realize that the Tesla one must be far more advanced and started asking around.

The Tesla robot will need to get out of an autonomous vehicle and figure out how to get a pizza up to your apartment, or to your front door, once you figure it out by talking to so many people, like I do.

Kaehler showed me a way how Giant’s AI would do that if it had access to the Tesla data and resources, particularly its simulator where rafts of people can “jump into” and walk around keep training over and over teaching it to get it right. The demos you see in the video are quaint compared to what the resources of Tesla can generate, as impressive as they are.

Every day I’m more and more convinced I’m conservative. Either way, getting the first look at Giant’s Universal Worker gives us a good look at the future of work so I hope you appreciate being first to see this. I sure did.