Subscribe
Share
Share
Embed
Deep conversations with the founders, investors, and operators building real-world AI - robotics, automation, industrial systems & AI infrastructure. Past the headlines, into how these technologies are really built, deployed, and scaled. Hosted by Bogdan Cristei, venture partner and former systems engineer.
THE OPTIM UPDATE
Why Pixel-Based Robot AI Is an Expensive Detour | Chao Cao, Sancho
Guest: Chao Cao, Co-Founder & CEO, Sancho
Host: Bogdan Cristei
---
Bogdan Cristei (00:00)
This is The OPTIM Update. I'm Bogdan Cristei, and today I'm talking with Chao Cao, Co-Founder and CEO of Sancho. Sancho builds the intelligence layer that lets general-purpose robots connect the machines in advanced manufacturing - the loading, unloading, and handoffs between process steps still done by hand today. Chao led the autonomy team for CMU's DARPA Subterranean Challenge and was a research scientist at the Boston Dynamics AI Institute.
He and his co-founder Jack quit their jobs in October. By December they had a paying customer. By March their work was on stage at NVIDIA's GTC keynote, and they just announced an oversubscribed seed round co-led by Fusion Fund and Catapult. We get into why $300,000 scientists are spending their days loading and unloading machines, why Chao thinks pixel-based robot AI is an expensive detour, and what three years of sending robots into tunnels and caves taught him about how brutal the real world is. Without further ado, here is my conversation with Chao.
Bogdan Cristei (01:10)
Hey Chao, how are you man? Really awesome to have you on today.
Chao Cao (01:13)
Hey Bogdan, great. Thank you for having me.
Bogdan Cristei (01:16)
Welcome in. You've been thinking about robotics for many years, so I'm super excited to talk to you. I want to cover a lot of ground today. To set it up: first I want to set the scene - why advanced manufacturing has automated individual process steps, but the connections between machines are still manual, and what that costs. Then I want to get into your technical thesis - why you think the dominant approach to robot intelligence is wrong and what you're doing differently. Then I want to hear the founding story, the customers, the GTC moment, the round you just announced - that's going to be exciting. And we'll close with some hot takes. Does that sound good?
Chao Cao (01:51)
Sounds good. Let's do it.
Bogdan Cristei (01:52)
Awesome man. Before we get into anything, I'd love to understand why you chose the name Sancho for the company.
Chao Cao (02:00)
Sancho is the name of the little sidekick of Don Quixote. He's loyal, down to earth, humble, and also witty. That's the character we want our robots to have while assisting human adventurers - hence the name Sancho.
Bogdan Cristei (02:15)
That's awesome. So help me understand the problem here. You make this point: advanced manufacturing has automated individual steps, and there's a gap between the machines. Walk me through what that gap looks like on the factory floor. In a typical advanced manufacturing facility, what's automated and what's still being done by hand?
Chao Cao (02:18)
I like to think about it as making lasagna, which is not so different from what advanced manufacturing does - whether that's cell therapy to cure cancer or AI chips at the end of the day. To make a lasagna, you need the grinder to grind the meat, the mixer to mix the sauce, and the oven to bake everything together. And there needs to be something or someone to connect those steps: bring the meat to the grinder, bring the sauce into the mixer, mix them together, bring everything into the oven, and take it out at the end.
At each step - the grinder, the mixer, the oven - those are the specialty machines we build to automate a single step or a single task. And we humans are extremely good at building those machines, however complex the step is. CNC cutting metal, 3D printing, even photolithography by ASML. However complex it is, we can build a machine for it. The challenge is at crossing the boundaries of those machines - bringing materials from one step to the next and making it seamless.
Traditionally we've only had two options. One is fixed automation: if you know you're going to produce a lot of product, you can build assembly lines and conveyor belt systems. The problem is the high upfront investment, and it's rigid. Only the biggest manufacturers, who know ahead of time how much they're going to make, can justify that cost. The rest of the world has to resort to the manual process - using humans as the universal connectors between machines. There's no alternative.
That's the third path we're building: use general-purpose robots to combine the flexibility of a human with the reliability of a machine, so we can democratize flexible automation. Not just the biggest manufacturers benefit - mid-size and even smaller manufacturers can have it too. Imagine a smaller group of people harnessing an entire floor of machines, so more product gets built per individual.
Bogdan Cristei (04:51)
Let's talk a little about that, because you say the most expensive labor is doing the lowest-value work. Walk us through what that math looks like. Who's doing the handoffs, what do they cost, and what happens when they make a mistake?
Chao Cao (05:01)
One example is one of our early clients - they do cell therapy. Scientists costing $300,000 a year are loading and unloading the machines. The machine itself can run at very high throughput, producing 150-something doses per day, but it's bottlenecked by the availability of the human scientist. And the contamination caused by humans is also huge, because cell therapy is a very expensive, very personalized drug. Human-caused contamination plus the throughput bottleneck is what drives up the cost of making this product. That's an example of where the scientist should be doing innovative scientific work rather than machine tending, material handling, and loading and unloading.
Bogdan Cristei (06:04)
You keep saying automation, and people hear automation. When I hear automation, sometimes I think: problem solved. Why isn't existing automation - conveyor belts, ceiling rails, industrial arms - the answer here?
Chao Cao (06:08)
When it comes down to a single step within the process, we can usually do a very good job at high-precision, high-throughput automation. The challenge is combining those steps in a flexible way - especially when future products will have much more variety. Different colors of a phone, different specs of an electric vehicle. You want the assembly line to stay flexible to meet the demands of different products, and fixed automation is not easy to adjust on the fly as demands and requirements change.
Bogdan Cristei (07:02)
Okay, now I kind of understand the gap. We can agree the dominant approach in robot AI right now is visual-language-action models - pixel-based systems trained on massive datasets. We were chatting before, and you mentioned that's the wrong path. So we get to your contrarian bet: 3D geometric world models instead of pixel-based approaches. Explain what that means in practice. What does your model see that a camera-based system doesn't?
Chao Cao (07:31)
At the end of the day, we're trying to answer one question: how resource-efficient can we get to the robustness needed for industrial use cases? Don't forget we're talking about industrial use cases, where the bar is five nines of success rate. How many resources do we need to get there?
People are talking about millions of hours of data collected for VLA approaches, and we don't think that's necessarily the right path. It comes from research momentum: since large language models, people figured out language, then moved toward image understanding with vision-language models, and then said, alright, let's put in the action piece, make it a VLA - vision-language-action model - and call it a day, hoping that solves robotics in the physical world.
From my experience deploying robots underground, and my co-founder's experience deploying autonomy on the most safety-critical application - on-road self-driving cars - this is not the case. The real world is much more brutal. It requires much higher precision for the robot to make contact with the world or change the world. We and robots operate in a three-dimensional world, and what matters is the precision you need to make contact with it. Geometry - the spatial understanding, the spatial structure - is what matters most a lot of the time.
Pixels encode a lot of information, which is needed for high-level understanding of a task. But when it comes down to safety, efficiency, and longer-horizon reasoning, it might not be the most efficient representation for the robot to capture what matters most - operating safely, efficiently, and robustly in the real world. That's the approach we're taking. And we're riding the tailwind of hardware getting commoditized - especially sensors. 3D sensing modalities are getting available and cheap enough to put on the robot and build this multi-modal representation, so the robot truly understands space rather than going through the detour of taking a picture and then acting.
Bogdan Cristei (09:46)
Got it. We also discussed your bet on test-time reasoning over brute-force data scaling. The analogy you used was AlphaGo and Chain-of-Thought in LLMs. Walk me through that. What does it mean for a robot to reason at test time?
Chao Cao (09:55)
This goes back a little to my experience in the DARPA Subterranean Challenge, where we sent robots underground to do search and rescue missions. One important challenge is that the robot has never seen the environment before. It has to figure out what to do and what not to do on the fly. That's the critical piece, because you can never rely on what we call the memorization paradigm - the robot has seen something before, knows what it did before, and reacts when it sees similar things.
The capability of figuring things out on the fly - that's what we call test-time reasoning. You reason about what actions you can take, how those actions will change the world, and then decide which actions lead to the best outcome. Because you will always run into situations you've never seen before. That's a critical lesson from my experience: you cannot rely on collecting a humongous training dataset and hoping that when deployed, the robot only sees things it has seen before. That's not going to work. The ability to make decisions on the fly, in real time, is super important to achieve the robustness needed for industrial use cases.
Bogdan Cristei (11:25)
Got it. You were mentioning that sensors are getting more capable and cheaper, so let's go back to you running your entire stack on a single onboard compute module. No external backpack, no cloud connection. Most demos I see require a bit more compute. How are you achieving that, and what does it mean for deployment?
Chao Cao (11:31)
This also goes back to our bet that a pure pixel-based representation is not the right way to go. Training on pixels and reasoning on images - especially with generative AI, if you want to dream the future in pixel space - is going to be very expensive. Pixels contain a lot of information, and predicting in pixel space requires a lot of compute, especially GPUs. And even if you get pixel-level accuracy, you're not guaranteed physical-level accuracy. Again, that's the detour we think is very expensive in compute and resources.
Going straight to 3D modeling - capturing what the robot needs to act in the 3D world, the geometry, the spatial structure - is a more compact representation of what matters for the robot to act. That lets us combine everything into a much smaller compute footprint and run everything onboard. It also speaks to my experience in DARPA SubT, where everything had to run onboard, and to Jack's experience running the whole stack on a computer inside the vehicle itself, without relying on the cloud or off-board compute.
Bogdan Cristei (13:11)
You mentioned DARPA SubT, which stands for the DARPA Subterranean Challenge. You spent three years of your PhD doing that, and you keep bringing it up, so let's have a discussion about it. What did the DARPA Subterranean Challenge teach you about building robots that work in unstructured environments, and what carried directly into what you're building now at Sancho?
Chao Cao (13:40)
You're right that I was literally in the dark for three years - spending my time in tunnels, caves, and subway stations. A quick intro to what the DARPA Subterranean Challenge is about: sending a fleet of autonomous robots into the most chaotic and messy underground environments on Earth to do search and rescue. There's a time window where the robots have to go in, find all the necessary information about survivors, and come back and report it. I was leading the autonomy team of the CMU entries, and I was the only human operator supervising the eleven robots we sent into the competition, because I was developing the algorithms and models for the robots to make decisions.
One important lesson from that experience is how brutal the real world can be. The competition wasn't even a real-world scenario, and it was already much further from lab settings, where you can make a lot of assumptions about perception and the environment - the ground is flat, we know the object shapes, we know the obstacles. Underground, it's GPS-denied. The robot has to figure out where it is with respect to the world, handle very rough terrain, determine where to go and where not to go, and coordinate between different kinds of robots - we used wheeled robots, legged robots, and drones all together. Those are some of the hardest challenges in robotics. It's a self-contained test of how autonomous robots can be in such environments - a good test of the frontier of current robotics technology.
Lesson number two is how to deal with uncertainty. No matter how prepared we were, there were always cases we ran into that we never anticipated. So it's ultra important for the robot to figure things out on the fly, in real time, rather than relying on prior knowledge or training datasets or things it has seen before. Those two lessons carried into building Sancho, where the robots again need to deal with dynamic, unstructured environments and handle unexpected situations. The most challenging problem in robotics is dealing with long-tail problems - handling all the corner cases. And the capability to deal with unexpected situations reduces the resources needed to tackle those long-tail problems.
Bogdan Cristei (16:15)
If I have my math right, you and Jack quit your jobs last October. By December you had a paying customer. By March your work was on stage at an NVIDIA keynote, at the GTC conference. Now you've just announced your seed round. I'd love to understand what happened from October to March. You quit, incorporated, and five months later you're doing interesting things. Walk me through that timeline.
Chao Cao (16:54)
It was an exciting half year. A little background: before this I was working at the Boston Dynamics AI Institute - now RAI, the Robotics and AI Institute - for a year and a half as a research scientist. But I thought the timing of the industry was mature for us to build robots. I'd say the industry timing is like the eighties and nineties for computers: hardware getting cheap enough and software getting good enough, especially with the advancements in AI. It was also good timing in our personal lives - Jack and I are young enough to be ambitious but old enough to be experienced, so we could start a company together.
We were exploring different ideas, and with some serendipity I ran into a former colleague of mine working at what is now our customer. They mentioned this loading and unloading problem, connecting the machines, and something clicked in our minds: this is exactly the kind of autonomy we were building before, and it could be put to use in these semi-structured and unstructured environments to bring value to advanced manufacturing.
Getting showcased in the GTC keynote opening was a validation point. It showed the possibility of end-to-end automation running on the floor, where general-purpose robots handle and operate specialty machines, and it becomes a seamless end-to-end workflow. That was a super exciting moment for us, and we've been building along this route since.
Bogdan Cristei (18:35)
Awesome. I think by the time this airs, you will have just announced an oversubscribed seed. Do you want to give us details around that?
Chao Cao (18:43)
Yeah, we're super excited that we just closed this oversubscribed seed round, co-led by Fusion Fund and Catapult, along with other amazing investors. Now we have the resources to build the robots and the intelligence and bring it into reality. Super excited about what's ahead of us.
Bogdan Cristei (19:01)
Awesome. One thing I didn't get to ask yet: your co-founder came from a very different part of the autonomy world. You come from research and DARPA challenges. He led mapping at Nuro and was a founding engineer at Phiar before Google acquired it. How do those two backgrounds complement each other, and where do they create productive tension?
Chao Cao (19:25)
Before the differences, there are actually a lot of commonalities we share. We met almost nine years ago during our CMU master's years. And the most important thing is we share this experience of deploying robots in extreme use cases. For me, it was deploying robots into the most chaotic and messy world on Earth - underground environments. For Jack, it was deploying autonomy into the most safety-critical application, self-driving cars. That taught us what it takes to put robots to work out there - the robustness and reliability needed to make it into a product, not just a fancy video or a lab demo. When we set off to build this company, we wanted to make something that works, products people can use, rather than fancy demos that live in labs. That's what we share.
The difference is I come from an academic background and Jack comes from industry. Robotics is still early in its life cycle as a technology, and we believe development needs to bring together state-of-the-art algorithms and models - the frontier of research - with the rigor of systems engineering. That's exactly what the two of us offer together.
And the tension is what you'd expect: I'm sometimes up in the air trying to see what's possible, and Jack is the one constantly reminding me what's tractable, so I don't lose sight of the practicality of making things work - and vice versa. It's a really good complement.
Bogdan Cristei (21:08)
Speaking of what's tractable and practical - I should say that when I see companies going into anything that touches health or bio, I get a little nervous, because I think you need a lot of experience to make a dent in that industry. But you and Jack decided to start in the highest-value, most regulated environments, and plan to work your way down from there. Most robotics companies I see do the opposite. I'd love for you to explain that logic.
Chao Cao (21:46)
At the end of the day, we're trying to find the intersection of what's technically possible and what's commercially needed - which is what everybody is trying to do. But we think about the difficulty of putting robots into environments a little differently. In these high-value environments - advanced manufacturing, biopharma, semiconductor, aerospace - the skills the robot needs are pretty similar to having a robot load and unload shelves in a grocery store, or do laundry folding in hotels. At the end of the day, it's about how much value we can bring to the process.
Maybe contrary to what other robotics companies are doing: bringing robots into a hotel or a grocery store is actually much more challenging than bringing them into more regulated, strict environments. Think about what the robot is doing. If a robot can make lasagna by connecting the mixer, the grinder, and the oven, then it can also connect different machines to do laundry at home. Those are essentially similar skills.
It's just that in advanced manufacturing, the environment is more structured. A grocery store is way more chaotic - you're dealing with day-to-day consumers - and there's less tolerance on cost for the reliability and robustness you need. So finding the sweet spot of market pull and the value delivered by the robots - that's where we think is a good place to start testing what the robot is capable of.
Bogdan Cristei (23:48)
Got it. Awesome. So I do have some questions I'd consider more in the hot take arena. Hot take number one: the VLA community says they just need more data and it will scale. You say it won't. Give us some of your flavor in thinking about why.
Chao Cao (24:09)
Let's say Waymo versus Tesla. The amount of data gathered by all the Tesla users driving cars around the globe is humongous - way more than all the Waymos combined, collecting data while running autonomously. And yet when you talk about end-to-end autonomous robotaxis, the approach Waymo is taking uses more sensing modalities combined, versus Tesla. That's the difference we're already seeing: the amount of data itself is not necessarily the determining factor. The quality of the data - knowing exactly what kind of data and what quality you need to collect - those are the more important questions to ask, rather than sheer scale.
Bogdan Cristei (24:53)
Got it. Next one: what's the most technically wrong thing robotics teams are still doing in 2026?
Chao Cao (25:01)
Maybe not technically wrong, but my general hot take: think about how much of the data collected today is going to be useless in the future. My take is that a huge, huge amount of the data we collect today is going to be wasted. At this point, we're still early in figuring out exactly what kind of data we need to make robots work - going back to our technical bet on geometry versus pixels. A lot of the time, people think: just collect whatever data is out there by strapping on some cameras, record whatever you have, and get the data scale up first before making things work. That hasn't gone through rigorous testing of whether it's the right path before throwing humongous resources at it. We don't believe that's the right way to do it.
Bogdan Cristei (25:58)
I record a lot of data too - I have my SO-101 arms - and it's amazing how many times I have to re-record the data. So it's interesting that you say that. Okay, third question and the last one: what do you wish you had known before starting a robotics company?
Chao Cao (26:24)
What robotics has taught me is that robotics is system work. It's not hardware alone, not software alone - it's a combination of many different things. Then I got to do this startup and realized it's similar to robotics: it's also systematic work. Not just the tech, not just the business, not just the people, but the combination of all of it.
I wish I had spent more time getting to know different industries beforehand. Before this, I was a researcher, a builder, but I didn't spend much time understanding what the rest of the world is doing. This is such a multidisciplinary field. I wish I had a better understanding from going into the factories and living their lives, rather than staying in the lab minding my own business. Getting more exposure to how the world works, how things get built - that's what I wish I'd had more exposure to before starting this company.
Bogdan Cristei (27:30)
That makes sense. Before we wrap, I do want to discuss the fact that you're hiring. In a few words, what type of folks do you need? What are you looking for?
Chao Cao (27:42)
Like I said, this is system work. We're not building just a hardware or a software solution - this is robotics, combining everything together. So the people we like to work with have that mindset. First, we really like the experience of deploying robots into the real world, because then you have the mindset and appreciation of how brutal the real world is. It's much more than benchmarking an algorithm or training models on computers - getting a robot to do real work in the real world requires much more than that. And agency: not "I'm working on the hardware side, so software is not my business," or vice versa. Ownership of the project and of the system. Those are the characteristics we're looking for in the people we want to work with.
Bogdan Cristei (28:43)
Any specific positions or titles you need within the next few months?
Chao Cao (28:47)
More toward perception - building the understanding of the world, especially for robots. Self-driving cars have one type of perception, and VLMs are a different perception model. For robots, we need a data-efficient path to the representation robots need to act. It's a brand-new problem that not a lot of people have experience solving, and something new we're going to develop together. So definitely roles there. And manipulation - especially whole-body loco-manipulation. Before, the field was maybe a little too focused on tabletop setups or industrial-type robots bolted down in one spot. Now we're talking about robots that can move like humans, go anywhere, and do the things humans can do. Combining moving and manipulation - what we call mobile manipulation - is a critical piece of innovation we need to develop together.
Bogdan Cristei (29:50)
Awesome man. And with that, Chao - where can people learn more about Sancho? Where should people go?
Chao Cao (29:57)
Sancho.com is the place.
Bogdan Cristei (29:58)
All right, sancho.com. Super good to have you on man. Thank you so much for your time.
Chao Cao (30:02)
Thank you for having me.