SK Gupta, Co-founder of GrayMatter Robotics: On AI and Human-Centered Automation

This article was written in collaboration with Michaela Gordon. Check out Michaela’s other writings at her newsletter: The Early

SK Gupta is the co-founder and Chief Scientist at GrayMatter Robotics, a company focused on robotic automation solutions for high-mix manufacturing applications. GrayMatter Robotics' smart robotics cells are installed in many factories in the US serving aerospace & defense, specialty vehicles, marine & boats, metal fabrication, sports equipment, and furniture & sanitary ware sectors. These cells are helping companies to increase human productivity, reduce ergonomically challenging work, compress cycle times, expand capacity and improve sustainability.

Dr. Gupta received a Bachelor of Engineering (B.E.) degree in Mechanical Engineering from the University of Roorkee (currently known as the Indian Institute of Technology, Roorkee) in 1988. He received a Gold Medal for securing the first rank in his B.E. class (1988) and a Gold Medal for the best Engineering Design Project (1988). He received a Master of Technology (M. Tech.) in Production Engineering from the Indian Institute of Technology, Delhi, in 1989. He received a Ph.D. in Mechanical Engineering from the University of Maryland, College Park, in 1994.

Key Lessons/Highlights

Customer needs are the North Star.

When asked how he thinks about prioritizing what to build and when, SK says that GrayMatter is extremely customer-focused. The team keeps customers at the forefront when coming up with new features and determining their product roadmap.?

The key to driving adoption of automation is to keep it human-centered.

We’ve seen strong resistance from workers against adopting automations which would eliminate many good jobs. In fact, limiting automation was one of the key demands of the longshoremen’s union in their latest strike. The key to creating automations which have a good shot at becoming widely adopted is to use a human-centered approach. In GreyMatter Robotics’ case, the jobs which they are automating such as sanding, blasting, and grinding are all highly undesirable jobs. These jobs all have high rates of injury and high churn (75% workers quit before one year). By aligning automation with the best interests of workers, you can greatly increase likelihood of adoption.

The gap between digital and embodied AI solutions is often underestimated.

It is easy for founders from pure software backgrounds to underestimate the difficulty of building embodied AI solutions. With digital AI solutions, the relationship between the application-AI relationship is the only relationship you need to manage. With embodied AI solutions, there are two additional relationships to manage, the hardware-application relationship and the hardware-AI relationship. A consequence of this added complexity is that going from development to production becomes much harder, as the surface area of failure has increased three times and embodied applications often have much lower error tolerance. SK estimates that it is at least two orders of magnitude harder than it is with purely digital solutions.?

Many recently successful AI founders have been in the field much longer than you think.

Many founders who have achieved massive success in the recent AI wave have actually been in the field for a very long time. These founders have been working on these problems for a long time and are finally having their moment due to recent breakthroughs in the field. They are rarely opportunists who pivoted to AI only after the industry became the new hot thing. As a prime example, SK has been researching AI for manufacturing since 1989, long before even simple AI was used in manufacturing. This is not to say that it is too late to get interested in AI, but rather to make sure that your interest is fueled by more than hype. Jumping from hype cycle to hype cycle rarely affords you enough time to build something generational.

Key Quotes

“There are a lot of exciting things that you can do in academia. You can take on new problems, study them, make new contributions, and you have a lot of freedom to study the kind of problems you want to study.”

“Our first instinct was that we had great technology in the lab, so let's get people to use it. Initially, we thought we could directly deliver? our technology to end users, like large aerospace companies. But they said, ‘No, we cannot get it from your lab because our people cannot maintain it. We only buy technology from a company.’”

“It was clear to us that unless we formed a company and did it ourselves, this technology wasn't going to go to market anytime soon.”

“Find customers, understand what they really need, and go build that solution. If you try to do a technology push, it's unlikely you'll be successful.”

“You really need to understand what market needs are and what people are willing to pay for. Build that, and deliver it to the market. That way, you increase your success probability quite a bit. Being customer-focused, interacting with customers, and understanding their needs is key.”?

“Another challenge that people face, especially those who come from a purely software background, is that if software works in the lab or in an office setting, making it robust and rolling it out to customers is hard but can be done pretty fast. But when you are working in the embodied AI space, going from something that works in your lab to getting it working reliably in factories is at least two orders of magnitude harder than what it is in pure software applications. ”

“We've changed human life in a pretty short period of time by bringing this incredible power of computing, sensing, the internet, robots, and automation. Many of these have been truly inspiring.”

Full Interview:

We wanted to start by talking a bit about your background. When did your interest in robotics start?

Sure. My interest in robotics started when I was an undergraduate student in 1987. I was doing a summer internship with a company that was making trucks. It was a highly automated plant in India. They had all kinds of transfer lines and hard automations. At that time, people were beginning to discuss how robotics can offer more flexible automation and what can be done with that. I was actually studying one of these highly automated transfer lines. That's when I started thinking about computers, automation, and robots.

Most of your career has been in research and academia. What drew you to that field??

At that time in India, personal computers were just arriving. We were beginning to see the potential of what could be done, how you could automate manufacturing, and automate decision-making. I graduated with my bachelor's degree in 1988. It was clear to me that I had to go to graduate school if I really wanted to leverage this computer revolution and use it to do something meaningful. So then I did my master's and came to the US for my PhD. As you study more, you find out there are more challenging problems where you need to study more. After my PhD, I ended up going to the Carnegie Mellon University Robotics Institute. They were starting some exciting work with robotics-related automation in the sheet metal industry. Since I enjoyed my time researching there, I decided to continue with that career because there are a lot of exciting things that you can do in academia. You can take on new problems, study them, make new contributions, and you have a lot of freedom to study the kind of problems you want to study.??

You spent so much time in academia. What ultimately pushed you to make the jump to start GrayMatter??

Around 1989 when I was doing my master's thesis. I was studying how to build a system that could, given a 3D model of a part, figure out how to machine that part automatically. For a computer to be able to automatically figure out how to machine that part, you need AI. So I started deeply studying AI applied to robotics and manufacturing.

The manufacturing world is kind of split into two distinct domains. The first is mass production. Historically, mass production has been AI-free. Automation is primarily hardware automation where you hand-program the robot, and there's no AI. There's a lot of automation, but it's all physical work automation, no thinking automation.?

Then, there's a whole other manufacturing process called high-mix production. Cars and electronics are mass-produced, but airplanes, ships, boats, yachts, and machine tools are not mass-produced. You make them custom to order, sometimes in very few units. Surprisingly, there's a lot of automation in mass production and limited automation in high-mix.?

That was our goal at the time when I went to graduate school: Can AI be used to automate high-mix manufacturing? Can we deploy robots in a high-mix setting and do a lot of automation there??

If you start tracking the journey of AI, the first wave of AI, which I started studying in 1989, began in the 1950s. The first generation of AI was all symbolic AI, expert systems, rule-based reasoning, and such. Despite a lot of effort put in by researchers, AI just wasn't mature, and we didn't have enough computing power to do anything significant with it. We were able to build demo systems in the lab, but it wasn't something that you could take to a real manufacturing site and get it deployed in industry. When I was at Carnegie Mellon University as a postdoc and research scientist, we started studying the second wave of AI, which was machine learning. We were trying to build systems that could be fielded in industry. We were successful in partnering with some industries and getting our systems deployed there. But these were not robotic systems, these were more quality control, quality assurance kind of systems. Later on, I joined the faculty at the University of Maryland, and we continued to interact with industry, but the penetration of AI in industry was very, very limited. Then came around 2010-2011, when deep learning or neural networks started taking off. That's where we started seeing a significant boost in AI capabilities. We started seeing AI being deployed in industry, but not yet in the field of robotics.?

Why do you think that is? Is it because of the difficulty mapping the “thinking” to the actuation?

Let's talk about the ingredients used in traditional AI. First, you need sufficient compute power. Now, for sufficient compute power, you need both a very fast onboard processor and a very reliable cloud infrastructure. This wasn't in place until about probably 2013-14, where we saw sufficient advances in that. The second part was that if you want to do highly flexible automation, you also need this new AI paradigm called physics-informed AI, which is able to function with limited training data. You also needed relatively inexpensive sensors so that you can collect data and build the workspace model. In summary, you needed AI advances that were not data-hungry, advances in compute infrastructure, and the availability of highly reliable, affordable sensors. All these pieces started coming together around the 2014-15 time frame. That's when it became clear that robotics could advance, going from mass production to a more flexible kind of setting. The first successful application of AI which took advantage of these advances was material handling. Material handling is when you pick a part from one location? and move it to another. You're not making any change in the part; you're just picking, packing, sorting, and transporting.

So those people started using AI-powered robots, and doing interesting things?

Right. But that's a relatively easy problem because your job is just to pick up the part, not damage it, and place it. That was the first wave of AI in robotics, lots of exciting things were happening with many new startups entering the picture. However, my interest wasn't in material handling; my interest was in processing. How do you sand a part, polish a part, or 3D print a part? That's my interest. But the same conditions driving the revolution in material handling—sufficient compute power, recent advances in AI, and availability of sensing—were relevant to processing as well.?

We started looking at it around the 2015-16 timeframe and realized the conditions were right. All the work we had been doing in our lab could now be fielded by industry to see what could be done in the processing space. At that time, many industry players were demanding solutions. Our first instinct was that we had great technology in the lab, so let's get people to use it. Initially, we thought we could directly give our technology to end users, like large aerospace companies. But they said, "No, we cannot get it from your lab because our people cannot maintain it. We only buy technology from a company."

Taking this feedback, we approached the robot vendors, but they said, "No, we only sell hardware. We don't sell software. Software is done by system integrators.” So we started talking to system integrators. They said, "Our model doesn’t use AI. We build dedicated custom software systems for people and integrate systems, but we don't use AI because we don't know how this will work." System integrators are very traditional thinkers.

All three pillars of the ecosystem—the end user who wanted technology but didn't want it directly from the university, the system integrator who had no interest in advanced AI capabilities, and the robot vendors who wanted to stay away from software selling—said, "Fantastic idea, the market needs it, but we are not going to do it." So it was clear to us that unless we formed a company and did it ourselves, this technology wasn't going to go to market anytime soon.?

Do you view the industry as replacing human workers or improving human jobs??

From the very beginning of my career, I have focused on those types of jobs that humans simply do not want to do. If I offered you a job sanding for 10 hours a day for the rest of your life, I bet you would say, "Are you kidding me?" There's an enormous amount of work in the industry, such as sanding, polishing, buffing, grinding, and blasting, which nobody wants to do. If you don't have any alternative, then some human has to step up and do it. But labor churn is extremely ? high in these segments. There's a 75% labor churn; people start at the beginning of the year, and 75 out of 100 people quit before the year is over. This leads to all kinds of challenges for the industry. You're constantly retraining and dealing with a new set of people all the time, which leads to quality problems and many of these tasks pose serious safety risks. In manufacturing processes, we focus on the kind of work that humans don't want to do anyway. The idea is to reduce the burden of ergonomically challenging work from humans and let robots do it. The human's job isbecomes to supervise the cell so they can optimize robot performance and cell performance. Both parties are happy. Businesses are happy because they're getting consistent, reliable results from the robotic cell, and humans are happy because they are doing more interesting work and not getting hurt. That's what we call human-centered automation. The goal is to deploy automation in such a way that the quality of life for humans becomes better. That's our philosophy.

Given the rapid pace of AI advancements in recent years, how do you prioritize which new advancements to start implementing at GrayMatter?

We are very customer-focused. We principally care about what our customer needs are and what features they want. We prioritize by determining which features they want, so we develop and deliver those features to them. If a new tool or technology comes into the market that appears helpful in delivering that feature, we'll take a look at it. However, we won't incorporate technology just for the sake of it. We have our own roadmap for rolling out different products and features, and we follow that. Whatever is appropriate in the timeline, we integrate it. There are a lot of advances happening. AI is divided into two different kinds: digital AI and embodied AI. Digital AI includes things like ChatGPT and image generation. The rapid rate of advances you see is all in digital AI. The second type, embodied AI, drives physical systems. In embodied AI, you don't have a new language model proposed every week. That's not what's happening. Embodied AI developments are not happening as rapidly as they do in the digital world because embodied AI requires an enormous amount of testing that sometimes takes months and years before any feature can be rolled out. The rate at which new things are announced is fundamentally different for embodied AI versus digital AI.

How did you meet your co-founders? How did you decide to start GrayMatter with them?

Both of my co-founders were my PhD students in my lab. We were working closely together and talking to industry at the same time. Both of them were super excited about all the technology they had built in my lab and felt that they needed to go out and deploy it. That's why we decided to form GrayMatter Robotics.

Got it. So you already had a very good sense of who they were, their capabilities, and all that.

With one of my co-founders, I had already worked with him because he was my PhD student, and he worked as a postdoc in my lab. So, I had known him for probably seven or eight years. The second co-founder was also my PhD student, and I had worked with him for about five years. So, I've known both of them for a significant period of time.?

Do you think you would have started a company with people with whom you maybe had less of a relationship, or do you feel that having that existing strong relationship is important before starting a company together?

I mean, I cannot predict what would have happened. You never know who you're going to meet and what is going to happen, but definitely having an existing relationship with people you understand well and who share your values and working style tremendously helps.

How has your experience in academic research and government research translated into startup work? Do you feel like it's a completely different environment, or are there any similarities?

Yeah, this is a completely different environment. When you work in academia, you are largely working on exciting research problems. Whatever is an unsolved problem, what is most intellectually stimulating and exciting, that's what you want to work on. When you're working in a startup, it has to be totally customer-focused. What the customer wants, when they want it–you have to deliver value to customers. Whether the problem is exciting or interesting, who cares? If the customer wants that particular feature delivered, you go deliver it. So, this is a very different world.

How have you enjoyed this new paradigm of work?

Remember, I'm still in a very unique position. I'm still a professor at USC. My day job is still being a university professor. I maintain my research lab at USC, advise PhD students, teach, and do research. The company, on a day-to-day basis, is run by my co-founders Bruel and Aryan. Bruel is the CTO, and Aryan is the CEO. I talk to them weekly and help them solve really challenging technical problems. I also advise on developing strategy and hiring the right kind of people for the company, but I still wear two hats. I get to be involved with the startup and academia through my work at USC.

What advice would you have for founders who also want to build at the intersection between robotics and AI?

So, advice is simple, right? Find customers, understand what they really need, and go build that solution. If you try to do a technology push, it's unlikely you'll be successful. You really need to understand what market needs are and what people are willing to pay for. Build that, and deliver it to the market. That way, you increase your success probability quite a bit. Being customer-focused, interacting with customers, and understanding their needs is the key.

Great, and then on the same topic of robotics and AI, what's one thing that you think is a misconception investors have about the industry?

I won't say misconception, but I would simply offer the following observation. Ultimately, you have AI, which is the software piece, and then you have a lot of hardware, which includes your robots and sensors and everything. Then you have the application, whether it is manufacturing, surgery, agriculture, or construction. Each of these applications also has a lot of know-how and constraints that get put on. To deploy a successful solution, you need to really get all three things packaged together: whatever application-specific knowledge tools are there, then the hardware and the software. I believe that sometimes people underestimate how long it takes to integrate all three things together. Some people say, "Okay, the hardware is already there, the knowledge needed for the job is well-understood, I'll just slap on software, and there we go." But unfortunately, it's not as simple as it may sound. You may have to spend a fair bit of time to get all these three things together, and all three are important. If you don't bring all three of them together, then you're not going to be able to get the performance and reliability that one would expect to deploy in the actual industry.

Now, another challenge that people face, especially those who come from a purely software background, is that if software works in the lab or in an office setting, making it robust and rolling it out to customers is hard but can be done pretty fast. But when you are working in the embodied AI space, going from something that works in your lab to getting it working reliably in factories is at least two orders of magnitude harder than what it is in pure software applications. All kinds of hardware problems will happen, and reliability problems will occur.

So, in software, we purely do digital work. If there's a bug, you can issue a bug fix a week later. You cannot do that style of deployment in a factory, right? You could cause serious damage to a part, or the robot itself could break, or someone could get hurt if you don't perform extensive testing. Getting that level of reliability, thinking it through, and making sure everything is safe and robust takes a lot of time and energy in the field of robotics. Going from a lab demo to actual deployment takes a lot of time. There's that element of physical safety that you don't have in software as much.?

How do you think being Indian has impacted your life or your founding journey?

I would say everybody is different, and everybody's journey is unique. I cannot decouple my personal experiences from what I learned as I was growing up. I cannot put on one hat and ignore everything else. Overall, my journey has been shaped by the time I lived in, namely when personal computers were getting popular and early AI was being pioneered. Coming to the US and studying in graduate school during that time has shaped my thinking. It would be difficult for me to attribute my journey to one specific component. Overall, this journey has shaped how I think about the world and what I'm able to contribute. I cannot isolate one piece from another because if I had started in the field 10 years later, my perspective would be very different. If I had started 10 years earlier, my life would have taken an entirely different trajectory. It's the overall timing, the people you meet who inspire you, and the specific set of circumstances that shape your thinking.

You mentioned people inspiring you. In the manufacturing, robotics, and AI world, who are the people you look up to, or maybe one person?

Naming one person would be an injustice to all the people who have inspired me. There are numerous people. People who taught me robotics, people who built incredible systems, and those I had the great fortune of personally interacting with and learning from. When I was studying at the University of Maryland for my PhD and then at Carnegie Mellon University, I met so many incredible people who have done great work.

So every interaction kind of inspires you and shapes your own thinking. Another source of inspiration for me is all the computing pioneers. The stories of Intel, Microsoft, Apple, Google, Amazon—all these are inspiring stories. Throughout my professional journey, I've seen Microsoft, Apple, Google, Amazon, Nvidia grow from their early days to what they have become today and the great value they have created. In some ways, all of those have been inspirational stories. We've changed human life in a pretty short period of time by bringing this incredible power of computing, sensing, the internet, robots, and automation. Many of these have been truly inspiring. Everything gives you a little piece in your puzzle, which then shapes your thinking and inspires you to contribute to society.?

Is there anything else you’d like to shout out??

I think the key message I would like to reiterate is that ultimately we all should be paying attention to improving the quality of life for our fellow humans. That's a good lens to evaluate whatever you do. The way I look at it is that all of us, the entire planet, aspires to have a good quality of life. Good quality of life comes when we can access products that are affordable, safe, and useful to us. We have to start viewing AI, robotics, and automation from that perspective. What role can these technologies play in uplifting the overall quality of life for all of humanity? I believe that if we develop the right kind of technologies and keep human interest in mind, we can improve the quality of life for all of us. That's how I think about it, that's how I analyze things, and that has been a useful perspective for me. I would encourage your readers to think about what we could be doing to improve the quality of life for all fellow humans.

Written by Edwin Ong and Michaela Gordon.