Bren Pierce on Why Humanoid Robots Are Overhyped and What Actually Works in Robotics

Transcript

[00:00:00] Bren Pierce: I didn't realize the CEO wanted a photo shoot. If they told me that upfront, I could've just gone there with a robot and charged $5,000. In terms of just cool factor, have they progressed faster or more slowly in three years? You're not gonna see that humanoid robot in your house and be like, make me a cup of coffee. You see that in videos, but that's because you've got 300 engineers just to make the robot take one step. If you push it too hard - nowadays you could do backflips and karate kicks - it's quite incredible. How do you persuade them to come to the startup when they can go and work at Google, work at Nvidia, or some of these big well-funded companies that are gonna pay you half a million dollars a year? It's quite hard to compete for talent there. It's scary but interesting when you land and you're literally like, I have nowhere to live. I have a stack of my stuff. You rent a car, you go and start living for four weeks in a motel, and if you don't have that DNA to actually get out of your house - I was forced to literally knock on the door of 50 hotels and say, can I speak to the manager please? To get it to be so generalistically reliable and just work - the classic 80/20. We're probably 80% towards it, but that last 20%, no one knows.

[00:01:16] Brian Heater: Hello everyone, welcome to Automated. I'm Brian Heater, the managing editor at the Association for Advancing Automation, and the guy who will be asking open-ended questions for the next 45 minutes or so. This week we are talking with Bren Pierce of Kinisi Robotics. Bren is also co-founder of Bear Robotics, whose systems might have brought you some ramen to your table at a busy restaurant. I think we actually had our first extended talk on the occasion of Kinisi winning the startup competition at last year's Automate. Bren will also be back at this year's show as part of our Humanoid Robot Forum. If you've been enjoying the show, it would mean a lot if you liked and subscribed to our weekly newsletter over at Automated.fm. And with that, please enjoy this conversation with Bren Pierce.

We talk a lot about what's coming up next in automation on the show, but if you really wanna see the future in motion, you've got to be there in person. Automate 2026 is where the world's leading innovators, builders, and dreamers come together to show you what's possible. Robots, AI, machine vision, motion control - you name it. All automation under one roof. And as part of Automate this year, the Humanoid Robot Forum brings together leaders, engineers, and researchers for a two-day deep dive into the real-world development, deployment, and commercialization of humanoid robotics. Register for free at automateshow.com to join us in Chicago, June 22nd through the 25th. We will see you there.

Does Kinisi identify as a US company or a UK company?

[00:02:58] Bren Pierce: Great question. It's more of a hybrid. We have our sales and marketing as a Delaware corporation in America, but I'd say 95% of all our engineering talent is actually based in the UK. Best of both worlds, one might say.

[00:03:15] Brian Heater: I feel like you should lean into the UK thing. It's a growing ecosystem right now - there are a few humanoid companies out of there that should be sort of part of your signature.

[00:03:25] Bren Pierce: I think being English helps as well. There's great talent here - you have Edinburgh, Cambridge, London. Top talent historically, but apart from Ocado, there just haven't been that many big robotics companies to come out of it. That's a mixture of the VC ecosystem - they're not really geared towards deep tech the way Silicon Valley is. What we've done is raise money in Silicon Valley and have our tech talent in the UK, which I think is the best of both worlds.

[00:03:58] Brian Heater: Looking through your CV, you've had a pretty diverse geographical journey of your own. You spent time at CMU, so you were out here - Pittsburgh is one of those places that probably identifies a little bit more Midwest than East Coast. Then the West Coast too - Bear Robotics for a number of years, and they're a Korean company - so you've had your hand in a lot of different robotics ecosystems.

[00:04:31] Bren Pierce: Historically the robotics community was actually quite small. You get Munich, you get Carnegie Mellon University, Bristol Robotics Lab, and places like that. You have KAIST in Korea. Up until about 15 years ago, I don't think there was that much robotics actually in the Bay Area. It's only about 10 years ago when the ROS community and Willow Garage really got started that the Bay Area began to become a magnet for robotics.

My advice for founders: the Bay Area has such good talent. There are companies from Fetch and others that have actually built companies there. You have to sort of go there, meet the ecosystem. The ROS Foundation is based there and great companies have been built out of there - Bear Robotics was primarily based in the Bay Area. But one downside - and why Bear Robotics actually hired massively outside of it, primarily in Korea - is you're just competing for talent in the Bay Area, which means salaries are high and it's just harder to attract people. How do you persuade them to come to a startup when they can go and work at Google or Nvidia, or some of these big well-funded companies that are gonna pay them half a million dollars a year? So it's quite hard to compete for talent there. That's also why at Kinisi, we hired in the UK. I think there's only one humanoid company in London. OpenAI is really the big company there actually hiring the next wave of foundational AI graduates.

[00:06:08] Brian Heater: Bear Robotics in particular really does have that Bay Area garage startup feel, connected to the ROS community - really going out, buying a TurtleBot, and building a company on the back of this robot.

[00:06:27] Bren Pierce: A hundred percent. I was part of that originally. My two co-founders Fangwei and John, both ex-Googlers and incredibly good computer scientists, went and bought a TurtleBot within a couple of days. They're like, alright, now what do we do? That's when I started chatting with them and we actually built Bear Robotics out of my co-founder John's restaurant. Me and Fangwei would be sat there coding during the day, and when the restaurant was serving customers in the evening, we'd actually run the robot to the customers.

Our very first product was literally just taking the bill and running it to someone's table and then bringing their credit card back to the till. That actually delighted the staff - they didn't have to walk to that table when someone waved at them, they could just put the bill on the TurtleBot and send it. Within about two or three weeks we had a whole system up that was just working. That gave us the confidence to say, alright, let's go and raise some money, start building custom hardware, and start commercializing this product.

John, the CEO, had a restaurant. He realized firsthand it's actually very hard to hire staff - they can be quite unreliable, people just don't turn up to work, and you're paying $20 an hour. He was like, how can I solve this? So he had the original idea to go buy a TurtleBot. But then I came in as the roboticist, with 10 years of experience with ROS, with the navigation stack, understanding occupancy grids and things like that, to actually help build out the robotics side. Originally he just thought about his own restaurant and then very soon he realized this is a problem a lot of people he knew in the restaurant industry actually faced. That's where me and Fangwei came in to help him build out the whole system.

[00:08:37] Brian Heater: So you started in software and SaaS, then went over to CMU and started studying robotics. Is that fair?

[00:08:46] Bren Pierce: Yeah. All my life I've been building robots as a hobby - obsessed with Star Wars, watching it too many times. When I originally graduated, at least in England, there wasn't really a robotics industry. It didn't even cross my mind to take what was my hobby and commercialize it. So I went out, found customers, bootstrapped a SaaS company, primarily for call centers to automate their operations for charities and all sorts of things. Then about four or five years into that, Bristol University opened the first robotics lab - the Bristol Robotics Lab. I read about it in the paper and thought, wow, I can actually turn my hobby into a profession. I luckily sold my company to our biggest customer and went in. I spent two years and did a master's in robotics.

But when I graduated, around 2008, you looked at the landscape and there wasn't really a robotics industry in the world. There might be KUKA or ABB doing pick and place in factories, but my passion was humanoid robotics and control, and there just wasn't an industry to go and work in. So luckily I went to Carnegie Mellon University and worked with Chris Atkeson there, working on the Sarcos robot. I also worked with Samsung and Honda on early humanoid robots, and that allowed me to break into the professional space and actually get paid to do my obsession - robotics - for 20 years now, which I'm very grateful for.

[00:10:25] Brian Heater: I wonder if that's one of those stories that's obvious and easy to tell in hindsight, but at the time - robotics was your passion and something you'd been interested in your entire life - selling your company and going back to get your master's, that's a big risk you're taking.

[00:10:46] Bren Pierce: If you look at what I've done my whole life, everything is a risk. A great story - how I ended up at CMU - I graduated with a master's in robotics and built for my thesis a 30-centimeter, one-foot-high humanoid robot that I made walk. There just wasn't anything there to hire me, so I literally emailed all the professors around the world that were doing something interesting and said, I will come and work for you for six months for free. Chris Atkeson from CMU said, great, this sounds like a perfect opportunity. And this is where you have to take risks in life. We didn't even do a Skype interview. I just said in an email, yes, and he replied with, can you start on the 3rd of October? I said, do I have to do anything? He said his EA will organize it. Within two emails he's just like, come. Here's an address. I'll see you at literally this room in Pittsburgh in four weeks time. I'm like, great - I guess I'm moving to Pittsburgh.

[00:11:54] Brian Heater: What was that ramp-up process like, moving to a different country, halfway across the world like that?

[00:12:01] Bren Pierce: It's scary but interesting when you land and you're literally like, I have nowhere to live. I have a stack of my stuff. You rent a car, you go and start living for four weeks in a motel - a Motel 6 on the side of a freeway - and you're just like, alright.

[00:12:17] Brian Heater: So the EA didn't set up that part?

[00:12:20] Bren Pierce: No, that was me. I think she would've done it a bit better than just a random motel on the side of a road. And you're like, alright, new place. It is also scary. You're walking around the campus knocking on a door going, hello, I'm Bren, what do I do now? He's like, here's a PhD student, here's a desk, have fun. And here's - I think at the time - a million-dollar humanoid Sarcos robot. Do something cool.

[00:12:47] Brian Heater: And that's the good outcome. Because my big worry is I would do all this, I would land there, I would get the car, I would get the hotel room, I would knock on the door and he'd be like, who?

[00:12:59] Bren Pierce: That is the big scare. You're on the plane thinking, I've never even spoken to him on a video call. This was like 2008 - Skype was a thing but no one really did that, it was all over email. I should say: in the intermediate time, after the EA contact, I did use Carnegie Mellon's department to get me a visa, so the university was involved. But with Chris it was literally a couple of emails - I'll see you there. And when you land you're like, what do you want me to do, Chris? He's like, here's the Sarcos robot - do something cool with it. So I got to work with some great talent, write some controllers, and get it working. There was also a project with Honda and Sarcos where it was like, alright, here's some robot companies - get it to do something interesting.

[00:13:44] Brian Heater: Okay so Honda coming along gives you some parameters or a goal to work towards. But day one - somebody gives you the keys to a very expensive robot and says, do something cool with it. How do you begin that process?

[00:14:03] Bren Pierce: Luckily they don't exactly give you the keys to the expensive robot on day one. They have a nice simulation program, so you're just like, alright, let's ramp up - install the software, figure out how their software stack works, figure out the controller, and just start writing controllers. Back then it wasn't anything like Boston Dynamics doing backflips. It was very much: can I just get this robot to balance? So when you push it, it doesn't fall over - it's more like a two or three-link pendulum. I spent a year primarily working on putting a balancing ball on the floor, pushing the robots around, and really going from a toy humanoid robot that just has servo motors and position control, to a fancy robot with force and torque sensors where you can write a proper controller using all the force information available on these expensive robots and actually just get it to balance.

[00:15:00] Brian Heater: So cool by 2008 standards would have been, don't fall over.

[00:15:07] Bren Pierce: Exactly. Ben - I think he's now one of the heads of Boston Dynamics control - he was there for about three years and his PhD work was just: push it with a pole and make sure the robot makes one step forward. I think he spent about five or six years on his PhD just to make the robot take one step when you push it too hard. And nowadays you could do backflips and karate kicks. How control has evolved in the last 20 years is quite incredible.

[00:15:39] Brian Heater: When you look back on a time period - and this is interesting, as we're recording this at the tail end of CES, so obviously we saw the new Atlas, we saw a bunch of humanoid robots trotted out on stage - in terms of capabilities, in terms of just cool factor, have they progressed faster or more slowly in the long run than you expected them to in that time period?

[00:16:13] Bren Pierce: I think the last three years it's just been incredible. For about the first 15 years, it was just slow incrementalism - you're walking, Jonathan Hurst at Agility was putting springs in the legs, and a lot of my PhD work was on series actuators, all of this, and it was slowly incremental.

And then three years ago someone said, we're just gonna use reinforcement learning - we're gonna put a robot in simulation, have 20,000 of them spend the equivalent of a million years in simulation to learn to control. Suddenly that was a big unlock. Before then you had incredibly complex algorithms: 20 or 50 different sub-algorithms trying to work together. You have a balancer, you're trying to work out center of mass, inertia, all of this complexity just to balance and make one or two steps. Then suddenly you put that all into an AI system that has no real understanding of physics, but you just brute-force incredible controllers that can do backflips.

[00:17:23] Brian Heater: So it's not just a marketing thing, it's not just people's imagination - something has really been unlocked in the last few years.

[00:17:32] Bren Pierce: A hundred percent. I think you have to break it down. There's control - can you walk, can you move your arms? That is basically solved now with reinforcement learning. The next frontier is going to be manipulation. Because now these RL-learned policies have just managed to brute-force: if I put my arm here, I'm gonna fall over - you do that a million times, you learn, oh, if I put my arm here, I don't fall over. Whereas before, people would spend years trying to figure out the algorithm. Now you're just like, the robot can't do anything, let's put a supercomputer in there, spend a week, and it spits out a controller that just works and can do a backflip.

There are caveats though. When you see a Unitree robot, it's always on flat ground - it's very much flat-surface-forward. If you put it on a slope of five or ten degrees it's gonna fall over. It doesn't really have that knowledge unless you train it on all these different terrains. But that just comes down to brute-forcing again - you just learn on all the different terrains, and that's what they're doing.

[00:18:45] Brian Heater: It strikes me that a lot - if not all - of the industrial and home humanoid robot industry may or may not exist in the near term, but to a certain extent it hinges on there having to be a somewhat similar breakthrough happening in manipulation. A lot of people I've spoken to think it's an impossibility to really crack manipulation in, say, the next three years.

[00:19:27] Bren Pierce: I would probably agree. In three years you're not gonna see that humanoid robot in your house being like, make me a cup of coffee - it walks around, it knows what's inside your cupboard. You see that nowadays in videos, but that's because you've got 300 engineers in a fake kitchen. They've modeled everything - they haven't hard-coded it, but they're like: that cupboard there has a tea mug in it, and they're all the same tea mugs, they know the coffee machine. If you want to drop it into any house and have it just know - walk around, open all the cupboards, oh, now I know where the teacups are, now I know where the dishwasher, now I'm gonna search online to figure out how the dishwasher works - that is the dream. I think that's still probably five to ten years. And let's be realistic - if I say five years, it's probably ten years away.

[00:20:22] Brian Heater: I keep having this conversation because I keep thinking back to conversations I was having with Gill Pratt on stage like five years ago, and we kept saying five to ten, and five to ten - the nice thing about saying five to ten is you can always just move it back another five to ten. It's abstract enough.

[00:20:42] Bren Pierce: I think because maybe I've been in robotics too long - I remember about 11 or 12 years ago, a friend of mine who founded Apex.AI, they do autonomous car software, we went around Palo Alto completely autonomously, about four blocks, about ten years ago, and I was just blown away. I thought the company had about 40 people. I'd been in an autonomous car. I thought, wow, this is incredible - jump forward ten years and we're only just starting to get out of geofencing in San Francisco. So I think that is the same with robotics and the home - we can have a lot of engineers figure out one small problem and get it working, but then to get it to be so generalistically reliable, to just work - that would just take so much time to iron out all the corner cases. The classic 80/20: we're probably 80% towards it, but that last 20%, no one knows.

I'm not going to sit here and say I'm the world expert in AI, but I just came back two or three months ago from CoRL - the top AI conference for academics. There were about six or seven professors on stage, doing this for years, and they couldn't even settle on what technique we're going to need. Is it reinforcement learning? Imitation learning? They couldn't all decide as a group whether that's the solution or whether we need a new architecture. They all had a completely different idea of how to get to the promised land of AGI in robotics. If they can't decide, that shows we haven't really got a concrete idea of what or how to solve this problem.

[00:22:27] Brian Heater: And I think this is really being positioned to consumers as, this is happening right now, this is happening in the next year - and it's hard not to feel like that's ultimately doing a disservice to the industry at large by setting these completely unrealistic timelines.

[00:23:01] Bren Pierce: I think this is where I see the analog with autonomous cars. I remember about ten years ago one company had a cool video - no steering wheel, and they were saying things like, we are designing ball bearings for the wheels that last a million miles because AI is solved for autonomous cars. You saw the videos and they're driving around their test facility and it just works. You think, wow, this is the future. Then ten years later, that company sort of doesn't exist anymore. And you see all these autonomous trucks where even I, ten years ago, was like, this makes sense - trucks that go from LA to San Francisco on a highway, very easy. A driver would get out in LA, it takes the very straight 101, and when it comes near San Francisco a driver gets in to do the hard bit in the city. Ten years later, you're like, this makes sense, but no one's achieved it.

[00:24:11] Brian Heater: And if that's the parallel - and obviously these things are never one-to-one, but a lot of people are suggesting it is a good parallel and it's a lot of the same people in these cases - that means probably the lion's share of companies that we saw at CES, a few years down the road, aren't gonna exist or at least won't exist in the form they exist right now. When you're building a company, how do you build for the long term and make sure it outlasts that initial bubble?

[00:24:53] Bren Pierce: I think I did this well at Bear and I'm trying to do it now. Instead of betting on foundational models and hoping my robot will work in the house, I personally think household robotic companies are dead in the water unless you are LG or Samsung or Boston Dynamics - some big company with that money to just work on this for the next five or ten years.

It's very much more pragmatic to focus on warehouse and factory automation where you drive to something using classical occupancy map-based approaches - because that works - and then use imitation learning and RL to pick up objects, drive, and put them down. No one's really doing that at scale, and the technology actually exists there. We've had to put in a lot of different packages and methods, and now it's about thinking: how do you actually get a system that a company can set up themselves without requiring my engineers to spend four weeks on it? That's the key lesson I've learned - you have to have a system that your end user can actually teach, train, or configure without needing a PhD roboticist. That's where a lot of robotics companies fail. They build really good tech and then forget, oh no, a forklift driver has to actually operate this robot.

[00:26:34] Brian Heater: So to a certain extent the AMR problem - is there even an AMR problem? The AMR problem is solved, right? These are deployed at large scale, they're good at doing traffic, they work with a lot of them around, they can operate alongside people. So the unsolved, big challenge is dexterous manipulation - that's gotta be the difficult thing for you to deploy at scale.

[00:27:04] Bren Pierce: Exactly. And I think the big unlock for Kinisi - and I think there are some other companies going after this - you go to a warehouse doing drop shipping or distribution. They have miles of racking, very cluttered. There's a box there and a human walks to that box, gets a knife, opens it to get out an ink cartridge, and comes back to the pick area. You could in theory do a Geek+ or Kardex or something like that where you spend 50 million on an automated solution with totes and it's all automated. But in so many warehouses it's just racking and racking and racking of shelves, just with a label that says A12, A13, A14.

So the goal for my company is: how can you set up a system where the operator or the system says, go get the ink cartridges at A12. You go to the shelf, you pick up the object, or you open the box and take one object out, and bring it back. No one's really done that, and that is a really hard problem. Going back to the household: if no one could solve this simplistic task of driving to a shelf and bringing back an object, how are you really going to do a household robot where you say, go get me a cup of coffee, and it's not all hard-coded with waypoints and prior knowledge?

[00:28:32] Brian Heater: And one of the things you've done - I'm far from the first person to comment on this - is it's a path of much less resistance to not even have to worry about legs in order to get there. That just gets you straight to market with a wheel base.

[00:28:53] Bren Pierce: It's even better than that. Because I co-founded Bear Robotics, we actually have a partnership where I just buy their base. It comes with their software stack. So for our example of driving to shelf A12 and picking off an item and going from the pack area to the pick area - that's solved. I can just say, Bear base, take me to A12. They have beautiful cloud infrastructure for mapping, for all of that. So that's a solved problem. Then we can just focus on how to actually open a box and take an item off it.

For your listeners, the big problem is a two-factor one. A: it's just the manipulation - how to randomly pick things, how to take them out of boxes. If it was all in a tote, you just go there, open a tote, pick an item out - that's classic pick and place, not too hard. The hard part is: how do you open up a box that's got 20 items in it? You actually have to cut it open. And the second part is: a lot of these shelves aren't uniform - how do you know you're picking out the Hewlett-Packard print cartridge X, Y, Z? A human can just see it, read the box, verify it without spending years or months tying into a warehouse operating system. How do you recognize 60,000 SKUs and make sure you're picking an ink cartridge and not a USB drive or something else random?

[00:30:20] Brian Heater: So right now, as we're recording this - and I have to mention that Kinisi is a really young company, it was a young company when you first came on my radar, it was a young company when you won our startup battle at Automate - you do have systems, you are shipping systems. If one of your customers buys and deploys a system, what can that robot actually do on a factory floor right now?

[00:30:50] Bren Pierce: Very simple. We have conveyor belts - a part comes out at the end of the conveyor belt, our robot picks it up. We trade it in about one or two days of setup. It takes that item, drives, and puts it into racking. Very simple, nothing groundbreaking, but it shows the system just works.

The hardware has been the main focus over the last year and a half - getting a whole system working. It took until about three months ago before you could just press a button and the robot boots and everything works. So now, luckily for my software team, the hardware is near enough solved - you're not debugging driver issues. We have all the integration with the Bear base, we have the safety controller so the arm just does what you want, everything's calibrated. So now we can much more focus on how to train the robots. We use imitation learning - we have a nice teleoperation system where you can collect the data, send it up to AWS or Google Cloud, train a model, and deploy it. That has taken us about a year and a half to get all working. Now we can focus the team on: we could do imitation learning, we could do a bit of RL - how to actually start getting that flywheel and data collection so seamless that we can give it to our customers so they can collect the data themselves. That's the big unlock we want to spend the next six months working on.

[00:32:16] Brian Heater: I think what you're getting at is right now, the big value add for them is that these systems are out there effectively collecting a lot of data. I'm always curious - I understand why Mercedes, say, would pilot these robots, as it looks like we're looking towards the future. But right now when these robots aren't really outperforming humans, what's the sales pitch? What's the value add for them to buy one of your robots?

[00:32:54] Bren Pierce: Reliability, and it just works. A lot of the tasks we're talking to customers about are just boring. Do you want to stand at the end of a conveyor belt picking up a part, walking two meters, putting it in racking, for eight hours straight? No one wakes up in the morning and thinks, I'm looking forward to that job, and after a couple of months, they just don't turn up to work.

So they're looking at reliability, and cost is always a factor - our robot is about a third of the price. But to give your listeners a bit more perspective on reality: at the moment our system is probably about half the speed of a person. So we couldn't go into a high-production Mercedes line and compete there. It's very much for lower-volume tasks that people go and do for four or five hours a day and then go do something else. We can just do it all day.

But for us as a company it's two-factor: A, it proves up the technology, and B, it gets that internal ability - how to actually deploy it. One of the key learnings is a lot of robot companies get to a stage where they have five robots and a couple of pilots, but they end up in a situation where you need roboticists from their own team to always go and deploy. It's so hard to scale because you have to go into the codebase and write custom interfaces and custom customer code. We are literally at that stage now. So we're focused on: how do you actually scale it? How do you get it so a big customer like Mercedes can take our robot and their warehouse workers can set it up in a couple of hours, like they would with a UR5?

[00:34:37] Brian Heater: I mean, I completely understand what's in these pilots for you. But if the ROI isn't there and is maybe a little ways off, why not just wait until you get to a point when it is there?

[00:34:56] Bren Pierce: A lot of these companies want to be on the forefront. Some of our customers have like 600,000 employees. They don't want to wait for it to become a product they can just go and buy - they have to learn themselves where to deploy it, and they use this as a way to educate themselves. What is the state of the art? How is it useful? Where can we deploy it? They can also cut through some of the BS they get in the sales channel, because if you speak to Figure AI, Apptronik, 1X, any of these - they're like, yeah, we can solve all your problems. And I think this is why a lot of these bigger companies - you might say they're wasting money - but they're educating themselves. What is the capability? And then they can reassess internally what is the realistic timeline to deploy these robots.

Also, luckily for the robotics industry, companies have come to us and said the CEO has seen a Unitree robot doing a backflip and wants 10,000 humanoids deployed next year. Someone in their C-suite has to now go and buy humanoid robots because a CEO saw a backflip and said, I want 10,000 next year.

[00:36:34] Brian Heater: That is one of the most relatable things - and it's not necessarily humanoid robots, but I think anybody who works in a corporate environment can relate to that from an AI standpoint. You're seeing AI everywhere, AI is the next thing, so somebody in a corporate suite who doesn't even necessarily know what the value add is, but knows that in order to keep shareholders happy and future-proof the company, you have to embrace new technologies. A Kinisi robot is an easier way to get in on the ground floor with a humanoid robot, certainly compared to a Figure or an Apptronik.

[00:37:35] Bren Pierce: And I think this is also a lesson in painful lessons learned. As a founder or a robot company, you have to go into some of these deals with your eyes open. A lot of these companies just want it for PR, so they can do a LinkedIn post or say we are X, Y, Z company and we have humanoid robots. They go into it like, I can spend 50 to 100 thousand dollars on a proof-of-concept pilot with a humanoid company, take some photos with the CEO. And then the problem is, as the company, you just invested 200 or 300 thousand dollars, all your engineers, and you thought they said they were gonna buy a thousand robots. I didn't realize the CEO wanted a photo shoot - if they told me that upfront, I could've just gone there with a robot and charged $5,000.

You have to go in there with eyes open. There is a two-way relationship - it's not too bad as long as you know what you're getting into. You go to a multinational corporation, get a photo with a famous CEO, you can use that for fundraising. But you also have to pay attention to: are these companies that have real problems we can actually solve? You're not just in innovation hell where they actually want to solve a problem. If we do well and hit clear KPIs, realistically they're going to buy the robot.

[00:39:10] Brian Heater: If the end goal was just to get people to take pictures with your robot, like, that's a much easier path to making income - you could just start a photo booth company.

[00:39:24] Bren Pierce: I literally heard about this about two months ago - a big logistics company brought in a robot company just so the CEO could take a photo and post, we're deploying humanoids next year. And then you hear back channel, how is that POC going? And they're like, yeah, it was just for a photo shoot.

[00:39:41] Brian Heater: That's the really slimy side of things. But there's also kind of a middle ground - a company genuinely interested in these technologies, seeing them do backflips at CES, deploying them, and then realizing they can't get there - and that's why they're abandoning their contracts. And then years from now, we spent all that money and we're not going to go back to that.

[00:40:08] Bren Pierce: This is where you have to go in with expectations set. One of the things I think I've done well is we've prescreened a lot of companies. The way we work at Kinisi is we speak to these companies - we get so much inbound, we haven't really done any marketing, but everyone seems to want to do humanoid. We ask them: can you just send us five to ten videos of problems you have and we will prescreen it. About 90% of them we come back and say, we like the problem, we understand it, but realistically we can't do it for two years. Get back to us. Because a lot of it is so complex - we could probably do it, but it's gonna be such a custom installation for you that it'll cost us a million and you're gonna buy two robots.

[00:40:51] Brian Heater: I don't know how you managed this or even how intentional it was, but your timing was really perfect as far as getting on people's radar at a point when the pendulum was starting to swing back. Years ago I was having conversations about AMRs and thinking, the next step is we stick some manipulators on these and they actually go and do the picking themselves. And then it seems like we skipped a couple chapters and now we're talking about robots with two legs. People are starting to see how difficult that is - the pendulum swung back a little bit to, okay, 70 to 80% of the problems could be solved with a wheel base - and then a company like Kinisi, and other companies with wheel bases, or humanoid companies that have a modular approach, are pushing that as well. It seems like you came along right as people were starting to have that conversation again.

[00:41:57] Bren Pierce: It was luck, but you create your own luck. About two years ago I was at IROS - the humanoid and ICRA conference is the big academic conference - and I was just sitting in on AI talks. I've been doing this for 20 years, and suddenly I'm seeing a lot of these demos of imitation learning, diffusion action models, and I'm just blown away. Wow, AI could actually do stuff. Before, being in the industry, you'd kind of fake it - you'd put an X on the table and use hard-coded trajectories. Then suddenly you see like the Gemini demos, all of these things about two years ago. And you're like, wow, the robots actually work. The robots can actually do stuff.

I saw a very clear parallel to when I founded Bear - you have these open-source AI stacks written by two or three PhD students that sort of work. Great, they have pretty good functionality. What it takes is taking those open-source stacks, putting 20 engineers on them, and actually turning them into a product. It was also lucky that I'd founded an AMR company. I started Kinisi by sitting in Bear Robotics headquarters with a desk for free, borrowed an AMR base from Bear, and just started building a robot on top of it. If I didn't have Bear Robotics already and my PhD wasn't in legged locomotion, I think I would've just built a normal legged humanoid.

[00:43:35] Brian Heater: So you officially leave Bear around 2021, and Kinisi is really launched towards the end of 2023. In that period between, you're doing some consulting. When you left Bear, were you already looking for the next company, the next way to commercialize robots?

[00:44:11] Bren Pierce: To be fair, I'd been doing a PhD and multiple startups for about 20 years. I was like, I just need some time off. I actually went and spent six months in Whistler where I don't think I even wrote a line of code, just snowboarding every day. And you know what, it was really that impetus of those six months off - nothing but snowboarding, partying, and relaxing - that I basically got bored and then I was like, alright, I guess I should do stuff.

In that time I'd been going to IROS, and before I even founded Kinisi, I spent about three or four months coming back from one of those conferences just programming for the first time in years - no management, not working for a different company, just four months programming generative AI systems, learning how these foundational action diffusion policies worked, really retooling myself. Going from historically a control engineer - my PhD work was MPC, Jacobian control, and at Bear Robotics it was classical occupancy grids, A*, Dijkstra's, navigation - to suddenly: wow, you somehow magically give a foundational model data and it spits out control. How does this work? And then actually going through and understanding what the transformer actually is, how it works, how you can use it, and actually how to program a robot.

One of the earlier videos I released for Kinisi was getting ChatGPT to control a robot using MoveIt, using SAM, all these vision-based systems, so you could actually chat to the robot and ask it things. And then also just how to install OpenVLA and these sorts of software.

[00:46:13] Brian Heater: Well, once I get off the call with you, my producer Jana is gonna have a good pitch put together about why she needs six months to snowboard to take the podcast to its next level. So it sounds like Kinisi early on, for a certain period of time, was really just pretty much you.

[00:46:34] Bren Pierce: Yeah. Early on I basically saw these foundational models and wanted to explore them. I didn't even start with the idea of starting a new company. It was more like, wow, look at these new generative systems. This was also when ChatGPT came out. So I was like, let's try and learn how to use an LLM to actually program. And that just snowballed until I thought, I think I'm gonna buy a robot arm. And then six months later you're like, I guess I should just start a company. This was about the time I actually met my co-founder Ed. I've known him for a while and we just started chatting about the business case. He takes care of a lot of the business side - actually talking to customers, figuring out what they want.

[00:47:16] Brian Heater: Is there an advantage to almost backing into starting a company in that you have no expectations going into it?

[00:47:25] Bren Pierce: Yes and no. The advantage I think I had - instead of going and raising money and then figuring it out - is I had time to actually learn the technology myself. Me and Ed went out for about six months chatting with everyone from Wendy's to Walmart to anywhere, to figure out where you can actually use this new AI and robotics.

The original idea was actually to be a shelf-stacking robot. We ended up spending a lot of time at Walmart chatting to them about in-store inventory. But then you back into the idea - they're like, why are you starting in our stores? We put Walmart stores right in the middle of a city where there are loads of staff. Where we really have a problem is we put our distribution center in the middle of nowhere because it's cheap, but we can't get staff. And then we went round their warehouse, saw these problems, and I went to about 20 warehouses looking at the same situation: you walk in, there's no automation. Not even an AMR. No Geek+, no arms, zero - just lots of people. And you ask, why don't you have AMRs? They're like, it's too much money, we don't want to do the CapEx, how do we integrate it with our operating system because we all just use paper? They're like, if you just had a drop-in replacement where I could say, unload that van, put it there, do this, do that - we'd buy 50. How many warehouses do you have? 600. You're like, you'd take 3,000 of these humanoids? Yes. Okay - I think there's a business here. I guess I should found the company and raise some money.

[00:49:04] Brian Heater: Are current AMRs a brownfield solution? Like, they can just be dropped into a lot of existing facilities?

[00:49:14] Bren Pierce: AMRs just work. Because all you can do is give a waypoint - go to loading bay two - or an AI stack can actually just correct it and give it a velocity and a turn. You get the best of both worlds.

[00:49:35] Brian Heater: It's funny - I was reading an interview you did with The Robot Report a couple of years ago, and it sounded like you were still in that phase of really trying to figure out Kinisi's business model. Maybe it was the Bear influence - you were looking at Miso Robotics. You weren't straying too far from the restaurant side of things. But it's been interesting to watch a lot of these humanoid companies evolve. I'm thinking of Agility because they were so early, but we're sitting here during CES week and I was at the CES where they rolled out Digit in the back of a Ford. That was their initial attempt to commercialize. And it does feel like there are still a lot of solutions in search of problems happening in robotics.

[00:50:31] Bren Pierce: And this is something I never really talk about because it was, in a way, a failure. My first robotics company, Robotize, we did it 100% the wrong way. I just graduated my PhD - Technical University of Munich had an incubator where you take PhD students and commercialize. You literally sit in front of a whiteboard and they're like, cool, what did you do for the last six years? Let's make a robot. Just artificially sitting inside an office and coming up with a solution.

We were looking at Relay Robotics - they looked successful, let's just do what they do. So we're going to hotels, asking what they need, let's just make a room service robot. With a twist - we'll make a robot that goes to your room and is a minibar, so you can replace the minibar with a robot that comes to you. You go to talk to every single hotel, they're like, we love it, we'll buy it. So you spend two years building this. And then you suddenly realize the unit economics do not work. They're like, we'll pay you 1500 a month. You suddenly realize it takes four days to map the place, about two weeks to do the installation, you just spent 20,000 on installation, and it's gonna take a year and a half just to pay back the installation time.

And this is the problem where you just have a technology and go around randomly being like, let's just find a problem to solve, without actually spending a lot of time deep-diving into what is the business case, what is the development cycle, how many people is it going to take. There also seems to be a golden rule in robotics: you need about a hundred engineers. And people start very small. Like, we are going to automate drug discovery for pharmaceuticals - sounds great on paper, but there are only probably 20 companies in the world that want it. If you do it really well, you're gonna sell a thousand robots. Whereas you look at Bear Robotics or Locus Robotics - 25,000 units - and you have the same number of engineers.

[00:52:51] Brian Heater: We're running up against time. So - obviously not everybody has the luxury of clearing their mind with six months of snowboarding - but for people who maybe are recent grads, maybe getting their PhD, maybe at a big company, maybe unfortunately affected by the recent layoffs at places like Fetch or iRobot - for people looking to really start something and commercialize technology, where do you even start?

[00:53:34] Bren Pierce: I was actually talking to some people at the weekend about this. If you have the advantage that you could just get a normal job, get paid a good salary, you can build up some sort of proof of concept in your spare time in the evenings. Go and start talking to people. Go and see: do you even have that DNA to start cold calling? Go back to Kinisi - I would phone up random people at Walmart and be like, hello there, can I speak to you about robotics?

And if you don't have that DNA to actually get out of your house - I was forced to literally knock on the door of 50 hotels and say, can I speak to the manager please, to figure out what their problems were. If you don't have that drive, you either have to go and find a CEO or someone to do that for you. But you have to get outside your comfort zone.

You could take the leap and just jump in with both feet, or it could be better for most people to do it on evenings and weekends and see: do I still have that energy and drive at 8 PM after working all day? Because I give people advice for free, and I get tired at 8 PM after working all day in my startup - and then at 8 PM I have to be like, great, I've got about 30 email backlogs that I need to clear. So it's not like you could have a 9 to 5 and comfortably work on a startup evenings. It's nine till five and then in the evening you clear backlogs or speak to people in America or whatever else. So, yes.

[00:55:14] Brian Heater: It's so funny - I've been doing a podcast where I interview musicians and artists, and it's the same advice. Under a capitalist system, it's always: figure out how to live, and then you can start doing the other thing.

[00:55:32] Bren Pierce: Or, if you're lucky enough - I'm doing something wrong in life because I speak to people who are professors or PhDs coming out of UCL or Stanford that seem to raise 20 million by just saying, I have a PhD in AI and robotics at the moment. Maybe that's the approach if you can do it.

[00:55:48] Brian Heater: Absolutely. Well, Bren, always a pleasure. Thank you so much for taking the time.

[00:55:55] Bren Pierce: It's always great to be on your show.

[00:56:02] Brian Heater: Thanks so much for joining us for another episode of Automated. Thanks to Bren Pierce. If you enjoyed that conversation, you can check him out live at the Humanoid Robot Forum - that is June 23rd and 24th in Chicago, part of the Automate show. If you've been enjoying this show, the easiest way to support us is to like, subscribe, and leave a comment, or tell a friend. You'll find more info about this show and our relevant show notes along with our weekly newsletter over at Automated.fm. And with that, we will see you next week for another episode of Automated.