Name: Brains for Bots
Uploaded: 2025-12-10
Duration: 19 min 52 s

Investors are pouring billions into robotics amid the AI boom, but only a handful of startups will win in the famously capital-intensive and technically complex robotics space. Sequoia’s Stephanie Zhan and Skild AI’s Deepak Pathak unpack the race to build a general-purpose “robot brain,” separating what’s hype from what’s quietly transforming industry. Then, watch the future in motion as Skild AI’s robots learn, adapt, and perform live on stage.

Brains for Bots

Conferences

Stephanie, Deepak, thank you for being here. We're so excited to be here. Thank you for hosting us, Sally. I think we should start with the biggest central question right now. The AI boom has facilitated a lot of hype around robotics, but is robotics overhyped? That's a very good question and a very natural one. Like if you look at the whole history of AI right from 1954, from the days of Turing, robotics was the first application that AI was basically founded for, you can say. And if you look at all over the years, We all have seen impressive robotics videos for decades. Now, you may not have seen, but they are super impressive robotics videos even from the 1980s, 1970s, but still no robots around us, and it always feels like robotics is coming in the next 5 years, next 5 years, next 5 years. It has been 70 years. So is it overhyped? Is it not overhyped? It's a very genuine question. And one thing to look at here is the way robotics is done today, in the last, uh, I would say even 3 or 4 years. is very different the way it was done for the past 65 years before that. And that's a big change. Now, what is that big change? In one line, the change is Things in robotics used to be driven more by human intelligence. Like somebody super smart would look at this task X Y Z and have the robot preprogram the robot mathematically to do it. What has now changed is that these models or these robots can now learn from data. So that has shift from programming something to learning from experience. And that is what this new wave is about, which is so different about robotics. And this is, we know this approach generalizes from all the results we see around LLMs, VLMs, language models, and all that. And that is one big change from the all 5 year chunks before that and this 5 year chunk what we are seeing now, Stephanie. So to put that in perspective, Deepak has this wonderful saying that he likes to to repeat, which is that what looks hard is easy, but what looks easy is really hard. And I'm sure that all of you have seen the videos on X that go viral. It's a robot doing a backflip, a robot running in a straight line, a robot folding laundry or doing some kind of household chore, and all of that looks impressive in the abstract. You know, in a vacuum on X, it looks amazing. But it actually requires expert discernment. And I think that um this is what Deepak is getting at. What you don't actually see is whether this robot was trained for general intelligence. Whether it was able to do that task, actually interacting with that environment, or actually capable of executing that task if something in the environment were to change. A small tweak, for example. What Deepak and Skilled are capable of doing is that they're building this general intelligence that enables robots to do anything, anywhere on any robot hardware. And that's massively transformative. It's similar to what OpenAI and Anthropic have done. Um, in the world of digital intelligence, but instead they're doing it for the world of physical intelligence. And so, you know, the challenge I think for you as a consumer of all these videos is to really discern what is real and what's not. And even amongst the videos that go viral online, there is massive overhype, um, but there is. Real progress and real trajectory that we're working towards and that's what I'm so excited about with skilled. Well, this is the point both you and Deepak have been very, very clear with me about, which is that it's actually a lot easier to program a robot to do a backflip than it is to get them to climb stairs. You're smiling. Why? Please, I think we should maybe show that video. Let's do it. That, can we cue the parkour video? Yeah, no, that's, that's not yeah, here we go. The issue is I, I sent Ellie multiple videos to choose from with no clue of which one to choose, and she chose this one. Now this, now it will look super impressive. Wow, the robot is doing this. I cannot do this. Uh, and it, it is like. And it can keep on going. It's funny, right? But one thing you would notice here is There is very little interaction of the robot with the world. Apart from ground, the robot is not interacting with anything other than ground. Maybe there was one block in there to jump over that block by putting their hands over. Now if you think about this, so this going deeper into this, uh, OK, I'll let the video play. Nobody will listen to the voice. Well, I was gonna say the other thing about this too is, in part, a huge part of what you're doing is around a robot brain, right? What does that actually mean? Yeah, so one thing, uh, like, uh, if you think about this, this component, hard, easy, easy is hard, right? This was, this is not my quote. It's from Hans Moravec. He was also a CMU professor, uh. This was the culmination of 30 years of work from the founders of AI from 1950s to 1989, and this was their whole conclusion, heart disease is. Now what does that mean? It means that, for instance, uh, If somebody can do a backflip in humans or if somebody can throw a ball into a hoop from very far away, you would pay millions of dollars to watch those people on stage do that, right? But then if somebody goes and opens a door or climbs a stair, you would not be surprised. Why? Because every human can do that. Now if you think about both these tasks, what is different between the two? The difference is interaction with the world. Backflip is full control of your own body in a free space, right? Which is exactly what computers are good at. If you tell a computers everything about you, about everything controlled, they're good at doing that. One example is airplanes. We have had airplanes for almost 100 years at this point, right? They fly in the free space, so there's no interaction going on. But when you talk about picking up a bottle, a glass of water or going on stairs, what that involves is using your vision. To immediately to continuously correct where you keep your next feet on the stairs. Any one point here and there and you and you fall backward all the way. And that interaction is what is the root reason for human gender intelligence. It's called sensory motor common sense, which you don't appreciate because every human, almost every human has it. Yeah, I was gonna say when you put it that way, walking upstairs is an incredibly high. endeavor, um, I mean, Seth, when you're thinking about this, you know, Sequoia is famous for being interested in big markets. What do you see in terms of use cases, market opportunity? Well, I would highlight a couple of things, um, you know, if skill succeeds, we completely transform what robots are actually capable of today. The market's massive already, you know, you know of companies like ABB. Kanu, Kuka, Mitsubishi, Kawasaki, that alone, just on the industrial side, plus all the consumer folks, that's massive. But they're still constrained by having robots that are only built for specific things. And the real potential of what skill transforms is opening up the possibilities that are otherwise not possible for us to achieve. So if you think about Commercial settings, security and surveillance, robots always fail at stairs, elevators. But if we're capable of actually producing robots that can do all those things, we can tackle security and surveillance. Take industrial, all the different types of tasks that that would otherwise fall under edge cases that robots still fail at. If skill succeeds, they can tackle a whole host of industrial use cases that often pose dangers to humans because of the materials and cramped spaces that you now have to deal with. And then finally, in consumer settings, Um, you know, you see all these household robots, but they're only capable of doing one thing. It's your one vacuum cleaner, your one dishwasher, whatever it is. But if you're if we succeed at building the general intelligent robots, you finally have consumer robots that can actually tackle the whole host of household tasks that you now have. And so the opportunity here is, it's really all of the unlocked possibilities. The a parallel here is, you know, if you had thought about OpenAI and Anthropic back in 2016, it would have been. Hard to actually understand what the scope of market potential was because we were still living in this vertical AI world in software, but the real market for an open air anthropic is the market potential for all of digital knowledge work. And in the same way, that is what skills unlocks. It's the market of all of physical labor work, all of blue collar work today. And not only that, I mean that alone is transformative, but what's even more interesting is we can get all of this at accelerated pace and at a fraction of the cost. Because you have something that is generally intelligent, you no longer have to program these robots for specific things that take a long time to program. And they're only available at 6 or 7 figures today. But if they succeed, you now have this generally intelligent software that can run on any robot hardware for an order of magnitude smaller in cost. And so that's what's most transformative and excites us about skills. We're going to turn to the audience for questions soon, but one thing I wonder about is, um, especially with the influx of capital moving into this space, is it a winner take all market, a winner take most market? I'm curious for both of you. I think this is a very, very good question. Uh, it's always hard to predict about the future, but in, in my opinion, I would say that the answer is very technical. Why is it technical? Because in robotics, if you see, what's the main issue? The main issue is data. Like if you see the language model land, whenever these language models, whenever show some result, any company, let's say open, they release some result. You see 5 of the labs almost reproducing the same result in a matter of months or a few months. Why is that? Because data exists for language models on the internet, for vision models on the internet. Now, whether you can use it or not, it's a copyright issue, it's a separate issue altogether, but it exists. But for robots, there is no data on the internet. So how do you get, how do you begin? Then uh training all these models to start with. Like the huge success of AI is driven by big models, big data equal to success. And we are missing big data in robotics, which is where the key part is. Right now we have to use alternate sources to somehow bootstrap robotics to get it out there in the field. But once it goes out there in Whichever company gets out there first across huge diversity of tasks, that creates sort of a data flywheel from robotics, right? And the interesting part is that's different from self-driving though. Because in self-driving, people still drive cars because cars help them go from point A to point B. You don't drive a car to help self-driving. That's not your goal. But in robotics, so robots have to be useful for something to begin with. But once, if somebody deploys these robots in large numbers across large numbers of tasks, Then you create the data flywheel, and that becomes your mode in some way to go forward. And when that becomes your mode, it almost becomes more like rich gets richer. Like, the more data you have, the better your models are, and the more data you get. So it becomes harder to catch on. And this is a classic trajectory for markets. Which are not heavily democratized like it's monopoly duopoly, triopoly, I don't know. So the winners will really be the winners. Yeah, exactly. What's your take, Steph? I agree with that. I think that the market traditionally has been very segmented based on different industries that you specialize in, but the reason is that you specialize in making robots for automotive, for industrial, for something else. If skills is successful, then they completely create this new category above all the hardware players. They become the generally intelligent software that powers everything else, and so it has the potential to be more winner take all. And I would say that I deeply believe Skilled is the leading frontier lab for building all of general physical world intelligence. They're leading in pushing the frontier. This is just a snippet of what you've seen. Uh, there's much, much more that you should check out on there. There are some, there is a video that involves a chainsaw and the robot brain the legs off a robot that where the robot actually exhibits emergent intelligence and generalized intelligence, so you should go check that out. They're the leading, they're leading in pushing the frontier research. They're also leading in terms of actually getting the. Robots out with real customers in in the real world and then I was the third thing I would say is they're leading in terms of becoming the talent magnet for all of the most elite talent in this field. And so it's been really exciting just 2.5 years since our founding, and they're on an incredible trajectory and, and much, much more to come. I think that it's moving at an accelerated pace, um, and it's, it's exciting to see. I would say you're a little bullish. I would add 11 thing to the argument about uh this monopoly versus like uh uh uh multiple companies. What I meant this argument was more for the software, right? But if this is of the hardware, it's very rare that people would want to, everybody want to buy the same hardware. So hardware is very, very difficult, unless you have, unless like Nvidia, like some chip or something. But if you're, if you the end hardware, People just want preferences. If you, if your neighbor has some hardware, and your other three neighbors have the same hardware, you would want something more unique for yourself. And so hardware, in the hardware side, you would see multiple companies, because that's how it is. Multiple form factors, different designs, different uh uh appearances. What, what I made the argument about, uh, this data, that's more for the brain part. Kind of like phones with Android kind of situation. Questions. We have time for one question. Oh my gosh, there are so many questions. How do I pick? Whose hand was up first? OK, I'm trusting you. I'm trusting you over there. Um, Deepik, one of the things you mentioned was that every year in robotics, it feels like by next year or the following year, we'll have something that's actually more mainstream. What do you think is gonna be for consumer applications or for people in this room, the time period over which we'll actually start to engage with robots on a daily or weekly basis? Yep. Instead of giving time frame uh for consumer like that you can buy or not, I would say one thing before you see robots around you in houses, which almost seems like the, the main narrative these days. You must be, you must have to see robots around other areas like when you go to grocery stores, when you go to hospitals, when you go to uh other more structured hotels. We are not seeing that right now. I think first that wave has to come, and then once you are already familiar with robots around the world, you're seeing it in real life, then you will say, OK, I have to buy those things now for my personal use case, the same way how computers began. First, computers were used for professionals. They were more for companies. The idea of personal PC came much later once people got familiar with them around it. The same timeline you would see in the shorter term. What we will see next is More and more robots are deployed in the enterprise scenarios, which for the jobs, which are not where not everything is known or not pre-structured, where humans are involved. Let's say a human is doing some task on an assembly line, robot works alongside that. And that is not possible today. Today, robots that have to be only in cages and things like that. So you start from those areas, which is where, where skill is beginning as well. We are already deploying in a variety of areas, variety of, uh, Um, factory setups, but all those setups are the ones where there's randomness involved. And they act as bootstrapping, uh, uh, platform for us to go to the next stage, which is more unstructured, which is more like hotels, hospitals, things like that. And then they eventually stage is consumer. It's the same task everywhere, same skills everywhere. So it's, uh, the ordering has to be some somewhat like this. And I'm going to take it home. I'm sorry to all of you who had questions though, um. There is a lot of anxiety around robotics. As much as there are all the fun videos, there are also concerns about killer robots. There are concerns about robots taking, taking jobs. What can our robotics future look like? I'll start with you, and then we'll end with you. Well, I would say that you know there's a large labor shortage in a bunch of different areas. Um, there is incredible pain that people feel in just the churn of labor in different industries. So that's an area, both of those are areas where I think a lot of the services that robots will now be capable of offering will first start to tackle. It's the places where there are already pain, um, a high urgency, um, and desire to adopt fast. I would say there are 3 essays of robotics future. Three Ss, OK. The first S is safety. Uh, now, this is not robot safety. If you look at the human jobs today, right? We are right now sitting in this room, we are a bit privileged to be working here. That's our job. But if you look at, there are still so many jobs out there, where every day people go to work and they risk their life. Either for a short term basis, some risky jobs in a, in some scenario, or they receive their life from chronic issues from like continuously bending, you know, in a position and doing work. But people have to do this work to make the society function. So those are the jobs where you definitely need uh robots. It will only improve the quality of life of the workers who are working, doing these jobs. It's the first, safety of people are doing dangerous jobs. Then 3rd, 2nd, 2nd reason is If you think about jobs. While this is true, the narrative is that robots will take jobs, but if you look at the employment statistics right now, there are almost a million more job openings than people are available to fill them. At the same time, there are so many people who don't have jobs. What is the discrepancy? It's a discrepancy of It means that people want to do certain kind of jobs. So there's a huge, almost a million gap, and it's only increasing, uh, after COVID. It has really widened. So people want to do certain kinds of jobs but not other kinds of jobs, but they have to be done. Like today it's a huge labor shortage for blue collar, uh, workers for big factories, big, uh, big, big corporations, and it's totally the opposite narrative. a million jobs short, that's the highest, uh, we have seen, sorry, million labor short of the jobs available, right? Now the third aspect is, let's say we all, everything works well and robots do really well and they get deployed. Then the 3rd aspect is more of a task for all of us. It's a social aspect. It is true that when robots become really good, uh, in the future, there has to be some work around how do we handle the job aspect, the salary aspect, and what people do. And I believe there is a local minimum. It is a social issue, like, so that not only few people are benefiting from the same this evolution, everybody is benefiting because there's abundance of everything. But if we go beyond that, then there lies a scenario, a good scenario where everybody's doing things that they like, uh, work is more optional, and they are doing things that they enjoy instead of being forced to. And in that scenario, there are so many robots walking upstairs. Deepak Stephanie, thank you so much. Thank you, Ali.