Empiric Earth 總裁 Jen Vescio 訪談

Jen Vescio is President of Empiric Earth, overseeing product strategy, corporate development, and commercialization. She previously served as Chief Business Development Officer at Uber, where she led global business development and new partnerships in mobility and autonomy. Her experience also spans investing in and advising emerging technology companies, with a focus on helping physical AI move from research into commercial deployment.

Empiric Earth brings together Nexar and Nauto to provide real-world driving intelligence, safety models, and validation capabilities. The company combines billions of miles of driving history with records of rare events to help developers, fleets, and other organizations understand performance and risk. In this interview, Vescio discusses independent evaluation, responsible data governance, commercialization, and the infrastructure needed to scale physical AI.

You’ve had an extensive career spanning global business development at Uber, investments in emerging technologies, and advising companies in the Physical AI space. You initially joined Empiric Earth’s board in April before recently stepping into the President role. What attracted you to the company, and what ultimately convinced you to take on a more hands-on leadership position?

As an investor and advisor in physical AI, I kept seeing the same gap. Plenty of teams were running experiments, and very few had the real-world data to prove their systems worked outside a test environment. Empiric Earth already had it. More than 10 billion miles of real driving and 60 million edge cases is something you can’t buy and can’t fake.

That puts the company in an unusual position. We don’t need to compete with the companies building robots, vehicles or models. We can help all of them get to market faster by providing the real-world intelligence they need to manage risk, validate performance and build trust. From the board seat, I saw the opportunity to build something foundational to the entire Physical AI ecosystem.

What moved me into the President role was Zach and his vision. I’ve known him since we were both building partnerships in mobility, me at Uber and him at Lyft. When someone you’ve trusted that long brings you a hard problem and an asset nobody else has, you listen. We complement each other. He thinks about the category and where it’s going. I’ve spent my career turning new technology into businesses that scale. Physical AI is moving out of research and into real deployment right now, and I didn’t want to watch that from a board meeting.

Empiric Earth was formed through the merger of Nexar and Nauto, bringing together two companies with deep expertise in real-world driving intelligence and AI-powered safety. What unique capabilities does this combination unlock, and how do you envision the unified company differentiating itself in the rapidly evolving Physical AI market?

Nexar and Nauto were both founded in 2015 and spent more than a combined two decades measuring the same problem from two different perspectives. Together, we can understand what’s happening on the road and how drivers respond to it, giving us a much richer picture of real-world behavior and risk.

For the wider market, that’s the bridge between simulation and the real world. Most teams build and test in simulation, but the hard part is proving a system can perform reliably in real-world conditions. We can help partners discover scenarios their systems haven’t encountered, improve their models and validate performance against what actually happens. That can shorten development cycles, reduce risk and accelerate commercialization.

What sets us apart is our ability to enable independent evaluation across the Physical AI ecosystem. We’re not building the vehicles, robots or autonomous systems our customers are developing. We’re providing real-world intelligence that helps them understand performance, manage risk and build trust. That requires a commitment to neutrality, portability and interoperability. As we bring new applications, models and services to market, we want customers to be able to use them across different hardware and technology stacks, without being locked into any single provider. Our goal is to act as a “referee” for the entire ecosystem, not pick winners within it.

Physical AI is emerging as one of the next major frontiers in artificial intelligence, with applications spanning autonomous vehicles, robotics, and intelligent infrastructure. What do you see as the biggest obstacles preventing these systems from operating reliably in unpredictable real-world environments, and why is access to real-world data so critical to overcoming them?

The biggest obstacle is that the physical world was never written down. Language models got good fast because they trained on decades of text people had already produced. There’s no equivalent for a road, a warehouse floor or a sidewalk. Most of what happens there, like a near miss or a lane that closed this morning, never gets recorded at all.

That creates two problems. The first is that the physical world constantly changes. A system has to know what’s true now, and a photo from last spring can’t tell it about a road that changed this morning. You also can’t go back and collect yesterday. The second is fragmentation. No single company sees enough of the world on its own, so developers are running separate experiments with different datasets, models and technology stacks. That makes it harder to compare results, share what works and turn promising research into commercially viable products.

Real-world data is how you get past both. The rare events that break a system happen somewhere every day, and you only find them by observing a huge range of real conditions, continuously. In our US data there are about 19 near misses behind every collision, and almost none of them show up in a police report or an insurance claim. Zach put it well in a recent essay: safety is the study of things that did not happen. You can’t study them if nobody counted them.

Empiric Earth has accumulated more than 10 billion miles of driving history and over 60 million edge cases, with hundreds of millions of additional miles captured every month. How can this vast dataset help AI developers train and validate more reliable autonomous systems, particularly when it comes to rare or dangerous scenarios that traditional simulations struggle to reproduce?

Most systems handle ordinary driving well now. What still breaks them is the long tail: a crash, a pedestrian stepping out between parked cars, an animal in the road, a work zone that wasn’t there last week. Those moments are too rare for a test fleet to catch in useful numbers, and simulation can only reproduce what someone already thought to model.

Empiric Earth was built to find them. We don’t process every mile our network drives. We keep the small share that our models flag as unusual or risky, and that’s how 10 billion miles becomes 60 million edge cases a developer can actually use.

Teams use that record in three ways. They search it: a team working on construction zones can pull thousands of real examples of human drivers getting through them instead of scripting a few. They ask for what’s missing: if a system struggles with left turns at busy intersections, we can point the network at exactly that and collect new examples as they happen. And they check their simulation against reality. When a partner reconstructs real collisions and near misses involving pedestrians and cyclists, they can see whether their simulated world behaves like the real one before they ship.

That’s what makes this valuable beyond the data itself. Simulation is essential, but it needs to be grounded in what actually happens in the real world. Giving developers access to that evidence helps them identify gaps earlier, improve their systems faster and make better-informed decisions about when they’re ready to deploy. It also creates a foundation for more consistent, independent evaluation as the industry scales.

One of Empiric Earth’s defining principles is its independence from the autonomous systems and vehicle manufacturers it evaluates. Why is independent verification becoming increasingly important as Physical AI moves toward widespread deployment, and what standards should the industry adopt to establish meaningful benchmarks for AI safety?

As Physical AI moves from research environments into public spaces, trust becomes just as important as technical performance.

Today, much of the industry relies on developers to test and report on the performance of their own systems. That’s understandable at this stage, but as more autonomous systems enter everyday environments, customers, regulators and the public will increasingly want independent evidence that those systems perform reliably.

We’ve seen versions of this challenge in other industries. When you get into a car, there’s an established framework for licensing drivers, inspecting vehicles and defining expectations for safety. Physical AI will need its own approaches to demonstrating readiness and performance. For autonomous vehicles, that could look like a driver’s license: a standardized test showing that a system performs at or above the level of a human driver.

The question is how we establish meaningful benchmarks that reflect real-world conditions rather than relying exclusively on controlled tests or simulations. Those benchmarks need to be grounded in real-world data, account for rare and dangerous scenarios, and allow performance to be evaluated consistently across different systems.

That’s where Empiric Earth can play an important role. We don’t believe the companies developing autonomous systems should be the only ones evaluating their performance. Our role is to provide the independent real-world intelligence, models and validation capabilities that enable others to assess performance, manage risk and build trust. By remaining neutral and interoperable, we can support evaluation across different technologies and platforms rather than favoring any one approach.

Ultimately, establishing standards will require collaboration among developers, researchers, regulators and other stakeholders. We want to help provide the foundation that makes independent evaluation possible, so that trust is built on evidence rather than claims.

As increasingly sophisticated AI systems interact with the physical world, organizations must balance the need for comprehensive real-world data with privacy, security, and responsible data governance. How does Empiric Earth approach these challenges, and what safeguards are essential for building public trust in large-scale Physical AI infrastructure?

Trust starts with how data is collected, protected and used. At Empiric Earth, privacy is built into our approach from the beginning. We focus on capturing the real-world events that provide meaningful intelligence, rather than treating every moment of driving as equally valuable. De-identification and responsible data handling are fundamental to that process.

Data provenance matters just as much. As Physical AI moves toward commercial deployment, developers need to know where their data came from, how it was collected and what rights govern its use. Without that foundation, it’s difficult to establish the confidence needed to scale.

Different markets also have different requirements around privacy, consent and data residency. Our platform applies country-specific settings for each, so we can meet local rules without rebuilding the system. We have to account for those differences as we evaluate where and how to expand.

The broader principle is transparency. Anyone building infrastructure for Physical AI should be able to explain what data they collect, how they protect it and who can use it. Responsible data practices aren’t separate from building a successful business. They’re part of what makes widespread adoption possible.

Your experience at Uber involved turning emerging technologies and strategic partnerships into scalable businesses. In your new role, you’re responsible for Empiric Earth’s product strategy, corporate development, and commercialization. What lessons from your time at Uber are most applicable to scaling Physical AI, and how do you decide when to build capabilities internally versus pursuing acquisitions or partnerships?

Uber taught me that the right answer can change. We built our own autonomous program, ATG, because we believed we had to own the entire technology stack. Later we divested it to Aurora and came back into autonomy through partnerships. Both calls made sense when we made them at that moment in time. What changed was our criteria around speed, cost and risk.

So every build, buy or partner decision gets the same questions. What’s the risk? What resources and capital does it take? How fast does it get us to market, and what is the competition doing in the meantime? Does it fit the strategy?

Empiric Earth starts from a different place than most companies because we already have the hardest piece to build, which is the data. We build where that foundation gives us an advantage that’s difficult to replicate. Acquisitions make sense when they add complementary capabilities or accelerate our strategy. And partnerships can often be the fastest route to customers, new markets and commercial scale. Physical AI is too broad for any one company to build everything, so we want to work with the partners who can help make the entire ecosystem stronger and sequence our approach based on both the market conditions in alignment with our competitive advantage – our assets.

You’ve also been involved in the open-source AI ecosystem, including work with the OpenClaw community. How do you see open-source innovation contributing to the advancement of Physical AI, and can greater transparency and collaboration help accelerate adoption while maintaining the safety and governance standards these systems require?

Open source changed how fast AI moved in software. When more people can build on the same tools, research cycles get shorter and good ideas come from everywhere, not only the biggest labs. Physical AI is earlier on that curve. A lot of the field is still in research and demos, and the question now is how to turn that work into businesses that last.

I was partnerships lead at the OpenClaw Foundation, which launched to support Peter Steinberger’s vision for open-source, personal agentic AI. It’s one of the clearest examples I’ve seen of an open project speeding up real work, because people could take it, run it on their own systems and make it useful for what they actually do.

That’s the lesson I think physical AI should take seriously. Open tools let more teams build and experiment without starting from zero. Keeping data inside the customer’s own environment, instead of asking them to send it somewhere else, removes one of the biggest reasons companies hesitate. A fleet, a city or a manufacturer can put AI to work on its own data, under its own controls, and see value faster.

Together, those two ideas are how this industry gets from research and demos to products people pay for. Openness alone doesn’t make a system safe, so it still needs real testing, clear governance and accountability. But an open ecosystem where data stays with the people it belongs to gets more systems into real use, and that’s where the learning compounds.

At Empiric Earth, we’re interested in how open tools and data that stays in the customer’s environment can help the whole Physical AI ecosystem grow, while we protect the real-world data that makes our work valuable. We don’t believe advancing the ecosystem and building a sustainable business have to be competing goals.

You’ve spoken about the importance of bringing more women into AI leadership, particularly as the industry begins establishing the safety standards and governance frameworks that will shape how intelligent machines operate in society. How do you believe greater diversity in leadership can influence the development of safer, more responsible AI systems?

I’m careful with this one, because “add more women and AI gets safer” is its own kind of bias. The point is broader. The agents we’re building learn from what humans have already done and from how we teach them. If everyone shaping that teaching thinks the same way, the systems pick up the same blind spots.

Diverse leadership brings more perspectives into decisions about what an agent should value, including empathy, care and how a choice lands on the person on the other end. Plenty of AI risk is technical. Other risks come from systems making decisions that don’t reflect the values or needs of the people affected by them. Bringing different perspectives into those decisions helps us recognize blind spots before they become real-world problems.

What I want from these systems is simple. They should help people make better decisions and give us more time with each other in real life. Getting there takes balance and real controls from a lot of perspectives. No single group solves it alone and we need to motivate and encourage women to participate and help drive agent to human, agent to agent and human to human connections.

Looking ahead over the next three to five years, where do you see the greatest opportunities for Empiric Earth beyond its existing applications in road safety and fleet intelligence? As autonomous vehicles, robotics, and other Physical AI technologies become more integrated into everyday life, what role do you envision Empiric Earth playing in the infrastructure that supports this transition?

Road safety and fleets are where we started, and they remain core to our business. But the bigger opportunity is to support Physical AI by reducing risk, providing a validation standard and ultimately making the real world safer.

The fundamental challenges are similar whether you’re building an autonomous vehicle, a delivery robot or a retail intelligence system operating a real world advertising and commerce network. How do you understand what’s happening in the real world? How do you know whether a system can handle unexpected situations? And how do you evaluate performance and manage risk as those systems move into commercial deployment?

I see Empiric Earth becoming an important intelligence layer for the Physical AI ecosystem, providing real-world data, models, applications and services that help developers, partners and operators answer those questions. Neutrality, portability and interoperability will be important to that vision. We want our capabilities to work across different technologies and platforms, so customers can adopt the solutions that make the most sense for them.

We have an extraordinary foundation in real-world driving intelligence, and the opportunity now is to build on it through new capabilities, partnerships and commercial models.

Physical AI won’t scale because one company builds the best robot or the best model. It will scale when the broader ecosystem has the intelligence, tools and safety measures to put those systems to work in the real world. That’s the role we want Empiric Earth to play.

Thank you for the great interview, readers who wish to explore real-world intelligence and independent validation for physical AI should visit Empiric Earth.