A two-year-old startup in Irvine, California, just entered the unicorn club without ever shipping a single robot. FieldAI doesn’t sell hardware. It doesn’t assemble arms, tune actuators, or manufacture anything at all. What it sells is something far less visible — and if the thesis holds, far more valuable: the brain that sits on top of someone else’s hardware and decides where it goes, what it avoids, and whether the step it is about to take is worth taking.
From JPL to the jobsite
FieldAI was founded in 2023 by Dr. Ali Agha, a roboticist whose resume reads like an autonomy masterclass. He spent seven years at NASA’s Jet Propulsion Laboratory as a Technologist and Group Leader, leading Team CoSTAR — a consortium of JPL, MIT, Caltech, KAIST, and LTU that won the Urban phase of the DARPA Subterranean Challenge in 2020 and placed second in the Tunnel phase. CoSTAR stands for Collaborative SubTerranean Autonomous Resilient Robots, and the NeBula autonomy stack it produced became a reference point for how machines handle unmapped, unstable, GPS-denied environments.
At JPL, Agha was also principal investigator for DARPA RACER, the off-road high-speed autonomy program, and for the prototype Mars Helicopter–Rover coordinated autonomy project. Before JPL, he was a research staff engineer at Qualcomm Research, a postdoctoral researcher at MIT, and earned a PhD in Computer Science and Engineering from Texas A&M University.
That pedigree matters in a field crowded with founders who graduated from language-model labs. Most “robot brain” startups today are chasing the generalization story from the top down — take an internet-trained vision-language model and teach it to output motor commands. FieldAI is chasing it from the bottom up — take a decade of tunnels, dust, fog, mud, and near-catastrophic robot failures, and distill them into a model that knows what it doesn’t know.
One brain, every body
The pitch is radically simple: one model, every embodiment. FieldAI’s Field Foundation Models (FFMs) already run on quadrupeds, humanoids, wheeled robots, and passenger-scale vehicles. No hardware-specific fine-tuning. No pre-mapped environments. No GPS. The same core intelligence that guides a quadruped through a construction site in Santa Clara can, in principle, steer a humanoid through a warehouse in Osaka.
Under the hood, the stack is a three-part system. A Dynamics Foundation Model (DFM) handles how the robot behaves physically — detecting slips, stumbles, and recovering from instability in real time. A Multi-Agent Foundation Model (MFM) coordinates fleets of robots operating together in the same space. Around both sits a Safety and Risk Awareness layer that turns every decision into a probability-weighted calculation rather than a deterministic output.
Inference runs on the edge at sub-100ms latency. No cloud dependency. No connectivity requirement. The probabilistic reasoning, which the industry typically associates with slow Bayesian pipelines, has been compressed into something fast enough to drive a robot at walking or running speed.
FieldAI has stayed deliberately fabless. While Figure, 1X, and Agility build both the robot and its software, and while Boston Dynamics still owns every layer of its own stack, FieldAI positions itself as an infrastructure layer — closer to Android than to Apple. The bet is that hardware commoditizes and the brain captures the margin.
Proven on real jobsites
The customer roster is not public, but the deployments are beginning to leak into the open. In November 2025, FieldAI published a case study with DPR Construction that quantified what the stack does at scale. On an active jobsite in Santa Clara, California, a Boston Dynamics Spot running the FieldAI Brain autonomously collected more than 45,000 photos, walked over 100 miles, mapped four floors, documented 125,000 square feet of roofing, and scanned 500,000 square feet of interiors.
DPR superintendent Justin Schreiner, quoted in the case video, said the FieldAI system had been on site for roughly eighteen months, used primarily for tracking construction progress through photography, hazard detection, material movement monitoring, and overnight security scans. None of that work required a map, a pre-planned route, or a human operator watching a joystick.
Beyond construction, FieldAI says it has active deployments across energy, manufacturing, urban delivery, and inspection in Japan, Europe, and the United States — though specific customer names remain under NDA. Early partnerships with Japan’s construction sector and European energy companies give the company a foothold in two of the most safety-regulated regions in the world, a position that becomes more valuable as the EU AI Act and ISO safety standards begin demanding certified autonomy layers.
The IP and the headcount
FieldAI is beginning to wrap its technology in patents. Its first publicly disclosed filing — U.S. Patent Application 2025/0252306, filed February 5, 2025, and published August 7, 2025 — describes a machine-learning framework for terrain analysis that generates synthetic training data and expresses traversability as probability distributions rather than fixed values. Robots running the system predict terrain features such as slope, roughness, and step height, then reason about the likelihood that a step will succeed or fail.
The company has also been hiring aggressively. More than 100 employees have joined in the months leading up to the August 2025 raise, with a stated plan to double headcount again by the end of the year. R&D hiring leans toward locomotion and manipulation, signaling that the next capability frontier is complex dexterous work — not just navigation.
The stakes
If FieldAI is right about the layer hypothesis, the robotics industry will bifurcate. On one side, hardware manufacturers will race each other to commoditize chassis, actuators, and battery systems. On the other, a small number of brain providers will capture the software margin — the way Microsoft did to the PC makers, and Android did to the handset OEMs.
The $405 million FieldAI raised in 2025, from an investor roster that includes Bezos Expeditions, Temasek, Prysm, NVIDIA NVentures, Intel Capital, Khosla Ventures, BHP Ventures, Emerson Collective, Gates Frontier, and Samsung, buys several years of runway to find out whether that bifurcation actually happens. The competitive read is tight — Skild AI is at a $14 billion valuation, Physical Intelligence at $5.6 billion — but FieldAI has something neither of them can claim: a decade of DARPA and NASA data on what goes wrong when a robot gets it wrong in the real world.
No maps. No GPS. No hardware. Just one brain, everywhere.
