Ali Agha, founder of FieldAI and one of the minds behind NASA JPL’s robotic autonomy systems, recently gave a guest talk on his transition from research labs to the unforgiving reality of robotics in industrial settings. It wasn’t much of a technical talk though, more like a “glimpse” into the practical, often brutal, constraints of bringing AI into the real world.

“It’s not an engineering problem.”

One of his key provocations was that people in academia and even some corners of the industry treat real-world robotics as a purely engineering or data problem. But Agha emphasized that fixing all the engineering problems in these complex environments often leads to massively overfit, brittle systems — like “20 million parameters for one corner case,” which then fail when the domain shifts even slightly.

This really landed for me. I found his emphasis on operational AI — AI that works in the field, not in simulation or theory — to be refreshing. It echoed the need for AI to not just be correct, but robust, adaptable, and safe. Especially in robotics, where mistakes aren’t just statistical noise — they can be fatal.

However, my approach to problems are still quite traditional. I remeber attending Ken Goldberg’s Distinguished Lecture in December 5th, 2024 called “Is Data All You Need?” where he stresses the importance of “Good Old Fashioned Engineering”–I almost think (with what little experience I have) most problems can be boiled down to an engineering problem without giving FMs too much autonomy at this stage of our development.

Dynamic Foundation Models: More Than Just Parameters

While the details were understandably limited (startups do guard IP), he introduced the idea of Dynamic Foundation Models (DFMs) and Modular Foundation Models (MFMs) — ideas that, while structurally very different, reminded me of the layered reasoning in System 1 vs. System 2 thinking. Quick, reactive modules handling immediate tasks paired with slower, deliberative components for planning. Again, they’re not directly comparable — but the metaphor helped me grasp what FieldAI might be aiming for: a layered approach to AI that is both reactive and strategic.

I wish he had shared more about the architecture. But I also respect the boundary he maintained — after all, these are commercially driven technologies. What I took away more than technical details was the courage of the vision: a small team (FAIRI has just 7 people) trying to build open-source robotics models that can scale into real-world applications.

Industry vs. Academia

Perhaps the most interesting reflection this talk sparked for me was on the tension between academic research and industry R&D. Agha’s team, in some ways, seems like another generalized automation company, though one with a bold vision. It made me wonder: what does long-term, high-risk R&D really look like in startups that still have to deliver?

Personally, I still believe academia is the place for the purest form of research — slow, rigorous, and foundational. But talks like Agha’s show the value of people willing to translate ideas into action, even if the path is messy and uncertain.

If I could ask Ali Agha one more question, it would be: How do you strike the balance between modularity and generalization in real-world deployments, especially under the constraints of sparse data and zero-failure tolerance?

That tension — between general intelligence and grounded functionality — seems to be where the most exciting robotics work is happening.