Andra Keay outlines how recent advances in mobile robotics are accelerating AI progress. She begins by comparing Fetch Robotics' warehouse robots (Fetch and Freight) to the earlier PR2, highlighting a dramatic cost drop from $375k to an estimated $37.5k and a shift from outright sales to a robot‑as‑a‑service model integrated with warehouse systems. Keay then explains that the same sensor suites and navigation intelligence used in self‑driving cars are now being repurposed for robots, describing the rapid price decline of solid‑state LiDARs (e.g., Velodyne’s Puck dropping from $60k to $8k) and the typical sensor stack of autonomous vehicles (cameras, IR, radar, GPS, long‑range and short‑range LiDAR). She surveys the emerging startup ecosystem—Zoox, Drive.ai, Kindred AI, and others—emphasizing talent migration from big‑tech and even quantum‑computing firms into robotics. Keay argues that embodied cognition is crucial for artificial general intelligence, noting that warehouse robots serve as a semi‑structured testbed that bridges the gap to fully unstructured home environments. She showcases Savioke’s “Dash” as a well‑designed relay robot that uses minimal facial cues and simple sounds to convey emotion, avoiding anthropomorphic expectations that can hinder commercial adoption. Finally, she stresses safety and interaction design: robots must have safe edges, predictable behavior, and a minimalist aesthetic to operate reliably among unpredictable humans (children, intoxicated adults, crowds) in public corridors.
Chapters
This is Freight from Fetch Robotics, and I don't know— oh yeah, here. This is Fetch. Looks a bit like the PR2, doesn't it? Yeah. Okay, well this is the 5th, possibly 6th build post-PR2 that that team did. So really one of the most experienced teams in the world, and they iterated and iterated and iterated on the PR2 price point, $375,000. Fetch and Freight are not sold per se. It's quite a different arrangement. It's more robots as a service. It's about the integration into the warehouse systems. But I can tell you that that robot, if it was available for sale, and it is actually for research purposes, is probably more like $37,500, not $375,000. Wow. So 2 orders of magnitude improvement in 5 years. Those are the things that are moving robotics forward at a fast pace right now. We won't see too many iterations like that, but we're seeing enough. The first one was— don't have it— smartphones. The next one is self-driving vehicles, because every mobile robot needs a certain sensor suite and a certain intelligence package to navigate autonomously. It's the same systems that you need in self-driving vehicles. And right now there is a war in producing solid-state LIDARs that can be mass-produced, that are going to be cheap and affordable for robots, but they're doing it for the automotive industry. Where, okay, it's cutthroat consumer industry right now, isn't it? There's very little margin in an automobile. So if anyone remembers the first DARPA Robotics Grand Challenge for driving, those LiDAR sensors on the top of the car cost more than the car. Okay. There's easily $60,000 worth of sensors. 2 years ago, I think Velodyne released the Puck, the new latest greatest small LiDAR for vehicles. And it costs $8,000. OK? So your average self-driving vehicle has got between 4 and 6 stereo vision cameras. It's got It's got IR sensors. It may even have radar. It's got GPS. And it will have one long-range LiDAR, and most likely 2 to 4 other shorter-range LiDARs, again, like the cameras positioned at corners. So you're trading off detail in the closer range with the ability, especially at speed, to track a lot of what's coming. In further distance. So there is a real war on to produce cheap sensors. That's, you know, I'm interested in not just the robots, but the robotics technologies. And that one doesn't show off too many of the sensors. They've also been discontinued. That was one of those wonderful goals, which was, let's reshape the car. We don't need steering wheels anymore. So why not turn it into a passenger vehicle? vehicle, not a driving vehicle. And you see these trade-offs all the time because some car companies have made their whole reputation on being an experience for drivers, and they're still going after self-driving but as an assistive function. Whereas others have said, no one wants to drive, let's turn the car into the passenger vehicle. But that's been pushed back a little bit, at least by Google. There is no telling what some of the other companies who are building hardware as well will be doing. And I'm not just talking about Tesla. There are companies coming up in automotive that are really disruptive. They're new. We've got Zoox out at Stanford. We've got Drive.ai, for example. And has anyone heard of Andrew Ng? Pioneered computer vision, deep learning at Stanford, then went to Baidu. Well, he's now advising Drive.ai. There's a lot of heavy hitters that have stepped out of the large company roles to go into small startups that are doing driving. What to me is very exciting is that right here in Silicon Valley, you've also got people that were in companies like D-Wave doing quantum computing, getting millions of dollars of investment money, building a robot startup now. So they've gone from DeepMind or D-Wave to building a warehouse robot. So Kindred AI, for example, they say it starts with a warehouse robot, but we're going to end up with embodied cognition and artificial general intelligence. Because remember when I said the thing about robots is that really we're building mirrors of us, and that's both flawed and and potentially black. But it's also really exciting, because with robotics you have the opportunity to be moving forward some things that can aid with medicine, whether it's prosthetic or advances, that can aid with our understanding of how the brain works, or how the human system works. But in any case, some people believe that to build artificial general intelligence, It has to be embodied. So robotics is essential for the next stages, the next steps for AI. I'm one of those people, but I still think that's a bit of a way off. However, where better to start with than working in e-commerce? It's the semi-structured environment closest to the unstructured environment of each of our daily lives, our houses. You know, if you think of the myriad of stuff that we have in our house, Where does it all come from? Well, most of the time these days, it comes from Amazon, or possibly insert several other large retail companies now. So what you do with those companies as enterprise robotics is the step towards becoming the consumer and household robot. I mentioned Savioke's robot. This one was called Dash. More generally, they're called relays, and they deliver small items. And they do it with style, with panache, and absolute cute. They do it terribly safely. There are a lot of things incorporated into the design of this robot. And this is an example for me of good design. Okay, you don't see arms. You don't really see a face. You get a little bit of an illusion of a face. But this robot does not talk to you. Okay? This robot does the R2-D2 thing. It just kind of goes beep beep beep. It doesn't even do the smile. But we interpret it that. It does a little bit of a happy dance and a noise. And that's all it takes for us to interpret that robot as having emotions. And we are much more likely to reward it for having done a simple task and succeeded at it than we would be to reward a humanoid-looking robot for delivering a parcel. Because, you know, shoot, isn't it going to, you know, go get a real job? Once we put too many human features onto a robot, we escalate our expectations. And I think that is a huge trap when you're trying to commercialize something. That is still really, really stupid. And so the designers that are doing a good job right now design downwards. They remove features so that you can have a beautiful experience that's easy to understand, very transparent interactions, because this is now moving up and down the hallways in offices. I think it's working with FedEx now as well as in hotels. So you don't necessarily know who's going to be in that corridor with you. They might be a toddler with bare feet, so, you know, make sure you have very safe edges. You might have someone who's drunk or someone who— a whole group of people blocking a corridor. Many situations that you can't predict, so you have to have safe behaviors.
