EP. 14

More Robot Data Doesn't Mean Better Robots | Raphael Levy, Foxglove

Raphael Levy of Foxglove on robot data quality, finding early customers and why robotics teams need people who understand the work their machines will do.

Discuss this episode and meet other Swiss tech builders in the TechDrive WhatsApp group.

Key insights: executive summary

  • More data does not guarantee better learning. Raphael argues that teams need to inspect data quality and understand how recordings feed into model improvements. Collecting petabytes tells you little about whether that process works.
  • Research spinouts need commercial skills. Engineers coming out of a lab may have an excellent product and little experience reaching customers. Marketing, sales and distribution need attention alongside the technical work.
  • Early pilots help you find out what matters. Waiting for a perfect product can postpone useful customer feedback. Raphael encourages founders to build proofs of concept with customers and learn what works during those first projects.
  • Your market may be outside your usual circle. A company in an unfamiliar industry could have a costly, repetitive task that your technology can address. Raphael suggests attending industry events beyond robotics and speaking to people doing the work.
  • Zürich combines technical talent with commercial ambition. From the teams he works with, Raphael sees strong engineering and an appetite for building businesses. This is his perspective on the ecosystem, informed by his experience in Europe and the US.
  • Domain experience belongs in the robotics team. Someone who knows a warehouse, farm or construction site can spot problems an engineering team might miss. Raphael sees room for experienced roboticists, newcomers using newer tools and people who understand the application.

Collecting robot data only helps if the team can learn from it.

Bring customers into the process while there is still time to change the product.

Why doesn't more robot data automatically improve a model?

A robot can produce enormous amounts of information without giving its developers a clear picture of what needs to change. Raphael’s concern is whether teams can examine that information, judge its quality and connect it to the next round of training.

He describes companies still relying on tools built internally years earlier. People may have become used to their limitations without realising what they are missing. For Raphael, the useful question is how well a team understands and controls the path from recorded data to learning. The size of the dataset alone cannot answer it.

What changes when a research project becomes a robotics company?

Raphael values research that starts with an interesting technical question before anyone has worked out the business case. He sees research and industry supporting each other, with ideas, talent and funding moving between them.

The demands change when a team starts a company. A strong engineering background does not necessarily prepare founders to position a product, find buyers or build a sales process. Raphael suggests hiring people with those skills, collaborating with them or learning the work directly. Commercial development needs a place in the company’s plans from the start.

When should a robotics startup involve its first customers?

Raphael pushes back on spending years perfecting a product before trying to sell the idea. During that time, the team may be missing feedback that would change what it builds.

He sees early proofs of concept as a chance to develop alongside a customer. Those projects reveal what works, what falls short and where the next effort should go. His advice is to start those conversations early and use the pilots to improve, rather than treating perfection as a condition for meeting a buyer.

How do you find a market for a robot that has barely existed before?

When asked about market sizing for new kinds of robots, Raphael turns to exploration. The obvious applications may be only a small part of the picture. Founders can learn more by visiting unfamiliar industries and asking how people currently handle difficult or repetitive work.

He recommends going beyond robotics networking events. An overlooked niche may have a clear need and customers willing to pay, even if it attracts little attention in technology circles. His answer is an approach to finding demand through conversations and testing ideas, rather than a new formula for calculating TAM, SAM and SOM.

What does Raphael see in Zürich’s robotics ecosystem?

Having worked in the US and with teams across Europe, Raphael notices different attitudes towards growth and profitability. In his experience, Silicon Valley often gives companies more room to pursue growth before becoming profitable, while parts of Europe expect a return sooner.

He sees Zürich bringing together qualities he values in both: a strong robotics talent pool and people who take commercial growth seriously. These are observations from his work with teams, rather than a ranking of cities. The conversation also touches on the challenge of moving from an early proof of concept to a business that can scale.

Who should be on a team building physical AI?

New coding tools are making robotics more accessible, and Raphael welcomes people entering the field through different routes. He sees newcomers and engineers with years of formal training contributing to the same team.

The discussion then moves to expertise outside engineering. A person who has spent years working in agriculture, construction or logistics understands details that are difficult to pick up during a short site visit. That knowledge helps a team recognise the actual problem and judge whether a robot would be useful. Raphael’s argument is for bringing these people into the work alongside those building the machines.

Who is Raphael Levy?

Raphael Levy leads Foxglove’s European go-to-market work, helping existing users and introducing the platform to robotics teams across the region. In the interview, he describes starting his career in autonomous-driving simulation in San Francisco, then working with annotation and training data for physical AI. That experience shaped his interest in how teams use data to improve their systems. He also shares robotics news, hosts conversations with people in the field and supports founders thinking through their route to market.

Episode show notes

A robotics team can collect petabytes of data and still miss the information that would help its product improve.

Raphael Levy of Foxglove joins Justinas Mišeikis for a drive through Zürich about turning robot data into useful learning and getting technical products into customers’ hands.

The conversation moves from autonomous-driving simulation and data quality to the commercial challenges facing research spinouts. Raphael explains why he encourages early customer pilots, looks for opportunities outside familiar markets and wants robotics teams to include people who know the jobs their machines will perform.

More conversations from the Swiss tech ecosystem

The next drive, in your inbox.

A short note when a new episode drops: the guest, the hook, and the one question you won't want to miss. No spam.

Follow the Swiss tech ecosystem, one drive at a time

New in-car interviews with founders, investors and researchers building Zurich's deep tech, AI and robotics scene. Want to discuss this episode, ask follow-up questions and meet other builders?