Narrow Tasks to General Intelligence: The Future of Robotics | Dreamforce 2025

Well welcome everybody to this session. I'm very excited to be joined by these folks. We look, we're at an inflection point right now with robotics, right? And robots used to be very much a simple task, they were repetitive. And we're at this like real inflection point. A new era of robotic innovation around general intelligence. So I'm excited to be joined by, by two folks today. And so why don't we start with Lyndon. And so Lyndon, congratulations. So Lyndon's co founder of Dyno Robotics, soldier company to instacart, you know, as a caper AI and really you know, thinking about the whole full stack of robotics, right? All the way from the hardware through to the model. And then I'm also joined here by Deepak and Deepak is the co founder of skilled AI Professor, Carnegie Mellon and, and I hear it doesn't do that many interviews. So we were like excited to have this conversation. All right, so Linda, do you wanna start a little bit, tell us a little bit about Dyna and what you're trying to achieve.

Yeah, sure.

Nice meeting everyone. Yeah. So here at Dyna, we build general purpose robots. Particularly we'll focus on high performance robots that could be commercialized in the commercial environment. As you generally know, foundation models is one where it's starting to show signs of life for very, very difficult tasks that traditionally wasn't thought that was possible. Such as things in logistics, manufacturing, warehouses, hospitality and so forth. But what we focus on is to make sure that we're able, our robots are able to perform at a level that is able to execute comparable to human level. And that is a core area of research. So as you can see here, we have, you know, launch a number of objective reasoning various diverse, diverse tasks across various environments. And our sole goal is to bring foundation model embody AI into a broader commercial scale.

That's awesome. Well, thank you for joining us. And then I also have Deepak, you actually want a video too? This will give everyone context as well.

Yeah, sounds good. Site skill we are building a general brain for robots. Think of it as single model that can work across any robot, any task, any scenario. I'm sure all of you have lost AirPods so you can see how hard the task is. These models can adapt on the fly. Whether your robot breaks or you have a different robot altogether, or you make up some robot by designing like a Lego piece, any body, one robot. So it's an omnibodied omni task. Omni scenario. Whether it's an arm doing things on the table or a humanoid in your house cleaning up your stuff. The same brain can adapt across all these scenarios. It's fairly general. It's trained on lots and lots of data. It is agile and dexterous at the same time. What you're watching right now is the only result of a humanoid climbing on stairs in the world. You can go back and check. You will not find humanoids doing long set of stairs out there. So that's who we are, skilled.

That's awesome. Thank you. And I didn't introduce myself, but I'm Saxena Romano. I'm our EVP and GM of field service at Salesforce. And part of where Salesforce has come from, as in field service, we've been managing and maintaining robots as assets. And customers all know a lot of different hardware and machinery as assets. And so our evolution has been in the advent of also here with agent force. Now we orchestrate work in a different way. And so general intelligence is giving us an opportunity to really think differently about orchestrating work. But also, robotics have advanced quite significantly too, to be multifunctional. And so Linden, maybe start sharing a little bit. Like what do you think has gotten us to this point right now that has allowed sort of this acceleration of robots? Like how do we get here? What has been a couple of maybe marquee innovations that have got us here?

Yeah, so I mean, historically robots has been fairly heuristic, rule driven. Traditionally it started with really just programming this exact position of the world where robots needs to be. That's probably used in car manufacturing. And then as we move over into the now what we call traditional, more ML stack on computer vision, sensor fusion type of systems, where it doesn't really handle things commercial environments very well, is that when you start putting into new environments or on scene environments, it starts breaking. Where foundation model has really converged is that it has things that are traditionally thought that was impossible, like folding, like chopping, like logistics kitting and so forth is now made possible because foundation model is able to navigate out of these corner cases. Historically, with traditional ML, you have, you know, in a typical folding scenario where you're folding a shirt, for example, it has different millions of permutations and millions of states. And now this could be done with more of imitation learning through diffusion models and visual language action models. So that's what we're converging towards. But I think one of the bottlenecks for the industry remains that data is still quite lacking in the industry to allow a larger amount of generalization. It's not like language Models where there's a boundless amount of data that you could collect from the Internet, physical action data is still quite lacking. And that's one of the critical bottlenecks with industry.

Yeah. And then, you know, I think Deepak, like we talked a little bit like this is, this moment here is also. You're talking about embodied AI across multiple different types of robots. What's had to come on the foundation model side to accommodate this to happen.

So robotics is a very different field compared to all of the fields in AI. Right. If you go back 70 years, AI actually was inspired by robotics. All the early work is 70 years. But look where vision has come, language has come in 70 years. And where robotics has come, we have no robots around us in general doing anything useful today. So the shift we are seeing in robotics now is not an incremental shift. It's a completely step change from where, how robotics used to be done. Just few years ago, robotics used to be done where people would program every motion of robot. Like you would have to think of a mathematical equation as to how the arm should move. Then you fit every parameter by hand and then you move it. It's a complete departure from that. The way things we do in robotics today, we can learn from data. Like this whole idea of learning from data and learning the behavior is something which was non existent in robotics. So whatever I'm showing today is like even just before we release these results, nobody would believe that this is even possible, that you can have a brain across bodies because it's not something a human can imagine. We are single bodied people. But in case of robotics we don't have any data. And, and if there is no data, why should I pick and choose? Which robot do I use to collect data? I should use every piece of data I can find on the planet. And that's the reason why we are doing so general approach to robotics. And we see fast progress because of that.

And so you both mentioned data because I think that is sort of a little bit of a challenge. Like you're saying like obviously the LLMs are a very different structure where there is so much knowledge out there that you and text that you can kind of compute with. How do you think about scaling this level of data? And obviously there's sims and there's also like teleoperated and there's only so much you can kind of really get a robot to learn. Tell us about data acquisition in the snow.

Yeah, so unlike language models. So we draw comparison because they are the only epitome of AI success. We have seen, unlike language models, there is no data for robotics on the Internet. And when we say data for, for AI, it's not millions examples, not billions, it's in hundreds of trillion. Like OpenAI no longer releases how much they are training on, but the release at the GPT 3.5, and that was 100 trillion tokens. 100 trillion. Like that's the scale. Even if the whole earth goes about data collection mode, they take them centuries to reach, which means we have to look for alternate sources of data to bootstrap the robotics problem from ground up. So what we do to get all these results we showed here, which is very general scenario, we use two such sources to start with. One is human videos. Because in my opinion, humans are also robots. We are just biological robot. Instead of metallic joints, we have biological joints. We have same motor neurons, sensory neurons. We see how humans move. In this scenario, we learn from that knowledge and then we combine it with simulation. Because with videos, I have no idea of forces. If videos were enough, if watching was enough, all of us could play like Federer by watching his games, practice and put the time into it. And that comes from simulation. So these two things together make like a very big data source and they are infinite. There is no limit on that data. But the difficulty is it's not the actual robot data. So you have to still figure out how to do the transfer and all that. This is how we bootstrap. Then we had pre training, post training, but that's the core code thing behind it.

Linden, you want to add to that? Thank you. Because I think this is like, you know, you're coming at it from a different problem set. So I'm curious what your thoughts are here on the hardware side too.

Yeah, so I think on the data scale problem, you know, our thesis is, you know, somewhat similar to Deepak in that we need to find a way where we could drastically scale up the data set to achieve general purpose human data simulation, real world data, teleoperations data, which is something that we use a lot more on, are all good areas. I think one setback on the teleoperations data by nature is that it doesn't scale very well. And I think that's where a lot of the people in the industry are practically looking at alternative data sources from simulations. I think world model is another one that's worth mentioning in human videos.

And so talk about, let's, let's talk about world models. But I think, you know, it sort of puts us in this world of knowing the domain the space so that it can be commercialized. Like one of the things that I think we talk a lot about is like robots are still not in that you mentioned really in the public domain. And part of that is around just ensuring governance and structure and all the rest of this. But commercialization of very specific use cases. I mean, we see this in AI agents actually right now, right? You're AI agents that are really productive. They're really focused on a single task. Right. And a sort of certain domain area. How are you both thinking about like commercialized? Like what does, what are your customers, what are, you know, organizations coming to you about around commercialization when it comes to a vertical use case and a specific use case that can be functional? Go ahead, Lyndon.

Well, yeah, so I think our general thesis is that you first want to fully understand what are the data sets and model capabilities that is fully required in order to land commercial impact. Because I don't think just pushing the frontiers of generalization is definitely very appealing. But what are we specifically pushing for? Because what we realized that most of the times in commercialization, general, it's not something that a generalization data set is going to drastically help. Most of the cases when you're chasing for human level throughput, human level efficiency, it's very similar to human, right? Like humans, we are general purpose in a way. We could do everything. But then the thing is, if you want me to go on an iPhone assembly line to assemble iPhones, I'm still going to, I still need to spend a lot of time repeating the same task over and over before I get really, really profit efficient at it. And that's one of our thesis, which is in the process of iterating and assembling iPhones, for example, I find tricks and I find ways to drastically improve throughput and quality of output. And with that insight that is actually very critical to start commercializing. As we think about scaling up the data set to get to general purpose, that level of insight at the data set level is going to be drastically helpful. As we think about how do we produce high performance general purpose and body AI models.

It's funny because as you were saying that I thought of like the Malcolm Gladwell's 10,000 hours. Even as humans, we need our 10,000 hours of a skill that we have to really get to for really great accuracy. So then, you know, like what are some of those use cases that you're seeing? Like really functional that you can see? Like there's a lot of ROI that people are driving to, but there's Multi purpose. And I think that's the balance because we're not talking about a cleaning robot like in that. And they're sort of like there's these areas in, I think domains where you have verticalization that you could, you're seeing more traction than others.

Yeah. So, I mean, one of the more public use case that we have shown is folding traditionally very difficult to commercialize with traditional robotics because of millions of different states that it could be in. And it also provides a very fertile ground for us to really understand the research that goes behind performance. So understanding what does it take to commercialize on one particular vertical helps us transfer that into new verticals that we've been working on. And we're seeing a lot of success in that.

Deepak?

Yeah, I think one of the things to really think about that robotics is not a question of if, it's a question of when. And you can see it around us. Right. I moved here from Pittsburgh. In Pittsburgh, you have to fix my roof. I have to wait two months for a contractor appointment. I have to deal with leaky roof for two months. And that's the wait time for getting a contractor. This is happening everywhere. So this is not one area which robotics can address. It will have the biggest impact of any technology we have seen so far. But what is required to get there. And again, going back to language models, in language models, the biggest breakthrough that is there in language models is this whole idea of in context adaptation. Like, nobody is training language models for your personal scenario, personal setup, but they still are becoming personal agents. Why? Because they can absorb the information online in context and adapt it to the scenario. That's what makes them flexible. There has not been a sign of in context learning in robotics. And in fact, in the results we are showing for the first time ever, some example of in context learning where the robot can adapt on the fly to changing scenarios. Once we can get beyond that point. So this will be like a slow adoption across few scenarios, maybe like in semi structured setups like hospitals or warehouses around people. But to truly get to consumer scale like that, everybody's home. We have to get beyond this same type of hurdle that Lemm went through, this in context learning. These chatbots used to be super useless five years ago. Nobody. So the joke area to work on if you're working on chatbots. But suddenly with this incontinence operation now this is the coolest area ever to work on. And that's the main hurdle we are getting. We are already seeing signs of in robotics, at least in our Setup at scale.

I actually really like that analogy because I think everyone who knows what it's like to have designed a chatbot dialogue, the utterances and like just the innate dialogue flow you would have had to build. And what has transpired now with agents.

Give me.

Can you. You just elaborate a little bit? Because you said semi structured. I think that's a really important part because we're talking about this in the context of enterprise, right? Our enterprise customers here. Like, what is semi. Like what does semi structured look like for you?

So there is a two extreme. One is like these factory settings. If you go to any warehouse, you have these robots all behind a red tape area. You cannot go there, you cannot touch there. That's structured. That's more like a microwave in my opinion. Like when microwave is working, you cannot put your hand inside it. It's a contained system. Other extreme is like completely. And it's like maybe my bedroom, like completely messy. That changes every day. In between are things like what I call semi structured. Like for instance, hospitals. They are random because there are people inside there. But it's still. They are not as random as a house scenario. Right. Similarly, warehouse, but warehouse, where people are working. Where people are working, people create randomness, they create messiness. And they. Which means that robot cannot be the go to the same position every time to do same thing. Those are the scenarios where we see the short term progress first. And this is the need of the hour as well. Today people don't have issues that, okay, they don't have any labor. It's like there is labor shortage. Like they want 100 people, they have only five people or 10 people. And that's the scenario. So it's okay to have people in the scenario have randomness, but robotic robotics will have the biggest impact in the semi structure. In between things. Yes, they're not too random, not at all structured.

Right in between guardrails. It's basically creating some guardrails just like we are in enterprise agents. That's interesting. Okay, so we only have a few minutes left. So I want to hear like, what do you think? What's next, Lyndon, what's next in the world of robotics? What are we thinking? Like, what's your timeline? I think that's the question everyone asks me.

Well, from our perspective, we already started commercializing against foundation models and manipulation scenarios. So for us it's really pushing the additional scale on manipulation, commercializing additional use cases across suitable environments. And at the same time really working on generalization research such that it's able to Provide a wide such that we're able to potentially zero shot new tasks very very easily. I think in the previous conference in Quora we've shown our ability to zero shot in new environments. Having really understanding and really drastically bring down the amount of effort that it takes to learn new skill sets and commercialize it in a level of, you know, speed of output, the throughput of execution, the quality of execution and the robustness of execution that could be, that could make our robots a lot more comparable to humans. Is going to be a very key question to, to. To answer in a, in a commercial environment.

Yes, I saw Dr. Fei Fei's behavior challenge 1K challenge of tasks. Right. And I think it'll be interesting to see what people do right. The 10 activities, 50 tasks that robots can do. Okay.

Issue with Benjamin is as soon as you make it the get sold.

Yeah, I know. That's why she like set the bar and it's going to keep going.

That's the whole problem with like a using. For us the next step I think is two things. One is deployments. We are already deploying at scale. Since we are kind of hardware agnostic, we don't have to convince companies enterprise to deploy our robot. We just say okay, you have a machine, have a robot, put your brain on this. So we are able to move very fast. We're already deployed at scale in multiple scenarios in manufacturing, packing, in mobility. In fact if you go to LaGuardia Airport Terminal B you'll find a robot going around measuring air quality index. That's the brain behind we are the brain. We don't have the logo on this, this is our partner. But the brain behind the robot is ours. Even right now. That's the one side deployment. It's very useful because robotics is not like ChatGPT. It won't happen overnight. People have to and Lyndon has experience doing this already once before in the previous company. It's a very hard thing to deploy and get roi get feedback going right to start early. Second portion is we have not yet discovered how we'll reach the robotics AGI. Like the research part on that is very much open ended. Like a lot of things to be discovered on that side. Similar to language as well. We have good chatbots but still a lot of things to be learned. And you can see every year is a new idea like reasoning. Before that it was scaling and all this. Similarly in robotics this problem is still unknown. So we are spending a lot of time in R and D and research to longer term goal as well as on the deployment side, these are two things which we are focusing on today. And these are also our next. So there's nothing new in the next. They're scaling this.

They're on their way already. They're already in motion. Okay. All right, we'll end with a hot take. Quickly. What is an area of robotics area that will not be disrupted by robotics? We. Maybe that can be a different hot take for either of you.

What do you think? Plumbers and surgeons.

Plumbers and surgeons. Okay, you took mine. Surgery.

Yes.

Yeah, I was about to say plumbers.

Plumbers. Okay. Well, you know what, that works really nicely for field service because that is our user base. So, you know, I'm just going to wrap with. Look, echo Salesforce 2. This is a really burgeoning area. Obviously we want to think about where our place is in the agentic enterprise where we orchestrate work. We believe that we're managing and helping our humans. We're leveraging and elevating our humans with AI agents. And now our workforce is extended to include the robotic workforce. So it is a part of our strategy too, to continue to think about what robots in the enterprise are not just in the manufacturing back office, but in that semi structured world. And I am in Deepak's corner. It's not an if, but it is a when. We're in a group of people that believe that, but in the end that it's here. And so thank you so much for spending time. I'm sure these two would love to continue the conversation, enjoy the rest of Dreamforce.