Seven lessons from founders building robotics businesses set to scale

Many of the tools available to applied robotics founders today didn't exist a year or two ago — and the pace of progress is striking even to the people building it. Ury Zhilinsky spent nearly eight years as senior staff engineer at Waymo, watching the field transition slowly and painstakingly from rule-based hierarchical systems toward end-to-end AI. Then at Physical Intelligence, one of the most well-funded robotics labs in the world, he watched the same shift happen in months. "When we started, we had no idea we could build something that could fold laundry within six months," he says.

At Bessemer's first-ever Robotics Day in San Francisco in March 2026, Ury, who now serves as Founding Technical Member at Mind Robotics, joined Sandy Hefftz, founder and CEO of Bellboy Robotics, and Ted Stinson, former CEO of Covariant, for a panel discussion moderated by investor Alexandra Sukin. Between them, the three founders have spent time at some of the most advanced robotics organizations in the world, Amazon, Waymo, Covariant, and are now deploying robots in hotels, warehouses, and automotive plants.

Here they share the lessons they think every applied robotics founder should know, from how to pick a vertical to how to hire for the roles that don't have job descriptions yet.

1. Build where buyers are primed to embrace robotics

Picking a vertical in robotics requires a delicate balance of factors. But for Sandy Hefftz, CEO of Bellboy Robotics, one signal is most salient: "The best places to build are where there is enough pain for robotics to be welcome," she says. For her company, the hotel industry checks that box. Staffing pain is acute for many of her buyers, as hotels often face labor shortages. One of her customers operates a 600-room hotel that runs on 150 people per day just to turn rooms over. These persistent pain points make prospects much more willing to entertain novel technology to alleviate it.

Hotels made an ideal vertical for several other reasons. The environment is structured: rooms follow predictable layouts, lighting is controlled, and tasks are repeatable. And importantly, the tasks are transferable to many industries. After Bellboy’s robots perfected the task of laundry sorting, they began being approached by laundry companies wanting to buy their robots. Now the team is edging into venues, where the team is deploying robots to set up tables for events.

Ted Stinson, former CEO of Covariant, remembers having customer conversations in the early days of the company to determine the best use case for the technology the team was developing. Logistics looked like an attractive vertical since tasks are semi-structured and operate at a large scale. “What we didn't know at the time that turned out to make it a really strong vertical to start with was the degree of customer urgency,” he says.

2. Building full-stack offers the deepest customer proximity, but partnerships are becoming an increasingly viable alternative

In Covariant’s early days, the team believed they could remain a pure model company, partnering with traditional automation providers to reach customers while staying focused on the intelligence layer. While an elegant strategy in theory, in practice it created a barrier. "The velocity and the customer insight were materially limiting to our ability to build a company," says Ted. The team eventually decided to become a fully integrated robotic system company, owning the hardware, the deployment, the customer relationship.

However, Ted is reluctant to generalize his learning to apply to founders today. He points out that the ecosystem looks meaningfully different now than it did seven years ago. "I see more and more examples of more traditional robotics and integration companies who have the potential to be great partners," he says. While the full-stack path still gives you the deepest customer insight and the tightest feedback loop, forming a great partnership is now a genuine strategic option under the right circumstances.

3. Don't define your vertical or your stack too early

Some in the field expect vertical robotics companies to dominate the next decade. Others think consolidation is inevitable, that the most successful applied robotics businesses will eventually converge around a smaller set of applications. But, as Ury points out, both views assume a prerequisite that hasn't been met yet: hardware that is actually reliable enough to build a clean stack on top of. “There is no clear proven embodiment that does everything right yet,” says Ury. “You could claim that humanoids will be able to solve all the problems, but this is hard in practice.”

“To have a clean breakdown, where somebody building the foundations, somebody else builds just the brains, you would really need to make sure that the foundations are solid and they can solve the problem," says Ury. Until then, he argues that companies that own more of the stack and iterate on everything together will move faster than those trying to specialize into one layer too early.

Sandy's experience at Bellboy illustrates what this looks like on the ground. Bellboy started in hotels, moved into broader laundry use cases, and is now setting up tables for event venues. "Vertical is a big word,” she says, pointing out the difficulty of separating out industries from the underlying tasks. The thread connecting hotels and laundry services isn't the industry, but rather it's the task type and the context layer the robot needs to operate in.

The longer a robot operates in a specific environment, the more context it accumulates, and this context is difficult for competitors to replicate. Every building where Bellboy deploys robots generates a persistent, building-specific model that learns the habits of the people working alongside them and knows, for example, that elevator three is slower than elevator two. Sandy suggests that place-specific knowledge of this kind may transfer across use cases in a way raw task capability alone cannot.

4. The hardest role to fill in applied robotics doesn't have a job description yet

At most applied robotics companies, the instinct is to hire researchers and engineers and figure out the rest later. But what often gets missed with this approach is the role responsible for managing the gap between the lab and the real world. Crucially, this gap is where deployments actually succeed or fail.

"It's one thing to get your system to run in the lab," says Ted. "It's another thing to actually see it working in the real world and see all the things you didn't understand because they only happen once an hour or twice a week. But they become material things that your system has to interact with." Navigating that gap requires someone who can make judgment calls across research, product, and real-world constraints simultaneously. Most companies don't hire for that explicitly, but they feel the impact when they don't.

"We learned that this role needed to exist the hard way,” recalls Ted. “We challenged some of our most elite researchers. They didn't like doing four weeks straight at a distribution center in Louisville, Kentucky." This experience spawned a role that Ted now calls a field deployment engineer. The ideal hire for the role is technically capable without being a full researcher, has deep customer empathy, as well as the communication skills to funnel real-world insight back to the engineering org. "There isn’t yet an equivalent classical job description for that role," says Ted.

5. Domain depth and conviction are essential to fundraising efforts

Convincing investors that physical AI was approaching a ‘ChatGPT moment’ was genuinely hard a year and a half ago, according to Sandy. Bellboy's first round didn't come from a VC. Instead, it came from a backer who owned 170 hotels and lived with the problem every day. This gave the company its first deployment and its first proof point outside of a controlled environment. "Today, there is a market with a real need and we have proof we can deploy it, so there is a clear path to scale," she says.

But the more compelling fundraising asset today, Ury argues, is domain depth combined with genuine vision. "It all boils down to deeply understanding what you're building and being able to deploy, learn, understand, and learn from the real environment," he says.

Yet past deployment successes and domain depth alone aren’t enough. "Nowadays there is enough evidence that applied robotics is viable," says Ury. "Now people are looking for somebody who not only understands the use case but is also a dreamer." Ted echoes the importance of a strong founder vision: "Sandy was telling me she's had this vision for a decade, and thought about it all the time, and visualized what it was going to take."

6. Every robotics company is paying the cost of building on models that weren't designed for them

Most robotics companies today are building on models designed for language and vision, not physical deployment. "A lot of companies are building on top of existing things that have not necessarily been built for robotics," says Ury. "There's a lot of companies saying 'let's add data, fine tune it, let's pre-train it.' But it's slowing everybody down."

At Mind Robotics, the response is deliberate: collect egocentric data directly on Rivian's plant floor rather than rely on borrowed training data. "Our entire strategy is based on the ability to collect data on the Rivian plants," says Ury. "Building on top of that is the foundation, the main thing that can get us to relatively high performance."

The second bottleneck compounds the first. For particularly hard tasks, pushing beyond imitation learning into reinforcement learning requires building complicated infrastructure from scratch, since there is no established playbook. "I'm really looking for that recipe," says Ury. "Having something like that would benefit everybody."

7. The robotics companies that win the next decade may be the ones focused on building the best integrations

Today's factories and warehouses were built for humans, with robots retrofitted around existing infrastructure. Ted thinks that's about to change. "The way factories and warehouses operate will reflect an automation-first mindset more and more," he says.

The pieces are already coming together. "The hardware building blocks are quickly becoming like Lego, and there are some amazing companies building physical intelligence layers," says Ted. He compares the future he sees taking shape to the Star Wars cantina, a room full of wildly different entities, each doing their own thing, somehow coexisting.

"What I see missing is the ability to rapidly integrate and deploy mass-customized systems," he says. "We're a lot closer in our ability to do that today than we ever have been."

For more conversations and investor roadmaps like this, dive into Bessemer's Robotics and Physical AI library.