Robots learned to see and to talk. But drop one into a real kitchen, lab, or factory floor and it is still clumsy — it fumbles a bottle cap, misses a box latch, folds a towel into a mess. What robots have been missing was never intelligence, but a pair of hands that can reliably engage with the physical world.

That is where our conviction in Generalist begins — one of the very few teams that truly understands where robots get stuck, and can get them unstuck.

A bet with unusual taste

If coding is the horizontal capability that defines an LLM's ceiling, dexterity is its equivalent in robotics: get it general, reliable, and fast, and you hold the key to most physical tasks. Generalist saw this early, concentrating on dexterity, UMI, and large-scale real-world interaction data while much of the field chased humanoids and demos.

What we admire most is the discipline behind the bet. In co-founder Pete Florence's words, "goals are more powerful than methods" — Generalist anchors on a hard, measurable goal and assembles the methods around it. It also holds an almost obsessive belief in scaling, like the early Anthropic: instead of chasing every data recipe, it nails pre-training on UMI data first, then layers in real-robot data. That restraint is where a deep moat comes from.

Data is intelligence

Robot training has long lacked data that is both scalable and action-aligned: teleoperation is high-quality but slow, human video abundant but hard to use for control. UMI is Generalist's answer — a hardware-and-data co-design that turns first-person human manipulation into a substrate real enough for policy learning, pushing the scalability of real-world data up an order of magnitude.

From thesis to product

In six months, Generalist shipped two model generations. GEN-0 (Nov 2025) showed that robotics has its own scaling law: trained on roughly 270,000 hours of real interaction data, with a clear capability threshold near 7B parameters, it declared robotics' pretraining era. GEN-1 (Apr 2026), trained from scratch on about 500,000 hours, lifted average success rates from 64% to 99% and ran ~3× faster than prior SOTA — on only about an hour of robot data per task. Research judgment, converted into progress — twice.

Why now

The race hasn't converged — model architectures, data recipes, and hardware are all in flux — which is exactly why now is the moment to bet. Once any core variable converges, as LLMs showed, capital and talent rush to the leaders and valuations re-rate steeply. The downside is guarded, too: incumbents are already acquiring labs that pair research with systems depth, keeping a frontier team like this a scarce strategic asset even if the winning path shifts.

Its endgame is simple and vast: a brain for many bodies — get the physical-intelligence layer right, then extend it across thousands of embodiments to billions of robots. We don't think the frontier of robotics will be defined by any single dazzling demo, but by the compounding ability to keep doing the right things right. That is what moves us most about Generalist — and why we are proud to walk alongside them.

The team

From day one, Generalist put foundation models, manipulation research, and real robot systems in the same room: Pete Florence (ex-Google DeepMind; Dense Object Nets), Andy Zeng (ex-Google DeepMind; foundation models and handheld data collection), and Andrew Barry (ex-Boston Dynamics; Spot's arm). It was full-stack in capability from the founding layer — never a model-first, systems-later company.

Pedigree is the easy part to list; harder to convey is how impressive this team is up close. Pete and Andy were already star researchers at Google DeepMind, and what stood out across our conversations was the rare consistency of their taste — each has circled the same problem for years. That gravity keeps pulling talent in: a DeepMind lead on robot world models, the hardware leaders behind Google Glass and Oculus, one of OpenAI's earliest designers — with zero attrition. Peers call it, unprompted, one of the very best teams in the field — among the most impressive we have backed.

To learn more, visit generalistai.com.