Teaching robots to use their hands: announcing our investment in Generalist!

We've watched the same story play out across AI for the past several years: a technology improves gradually, then crosses a threshold where it goes from "interesting demo" to something people actually rely on. LLMs did it, image generation did it, and then agentic code generation did it. In each case, the underlying progress was continuous (more data, more compute, steady capability gains) but adoption wasn't. Adoption happened in a step function once the technology got reliable enough to trust.

We think dexterous robotic manipulation is next, and Generalist is the company we're betting on to get there. We're excited to announce our participation in their Series B!

Pete Florence, Andrew Barry, and Andy Zeng are among the most credible founders in robotics. Pete and Andy spent years at Google Brain and DeepMind, where they co-authored PaLM-E, widely considered the first vision-language-action model. The fact that they helped create the dominant paradigm in the field and then chose to build something architecturally different tells you their conviction is informed. Every researcher and roboticist we spoke with said some version of the same thing: these are among the best robotics researchers and practitioners in the world.

What gives us confidence in the trajectory is how Generalist collects data and trains their models. Rather than learning from internet video, which captures what manipulation looks like but not what it feels like, Generalist trains on data from humans physically performing tasks while wearing motion capture devices. That data contains the things that actually matter for reliable execution: contact trajectories, force patterns, how objects deform and slip and resist. It's harder to collect than scraping video from the web, but the researchers we spoke with during diligence were consistent in their view that it produces better models for real-world commercial tasks.

And every commercial deployment generates failure cases, recovery paths, and edge-case data that flows back into training. Competitors working primarily from research labs can't easily replicate that feedback loop. If the scaling curve holds, and early results suggest it does, Generalist's lead compounds with every deployment.

Generalist is already deployed with early partners across automated storage, lab automation, and contract manufacturing, with a pipeline spanning automotive, electronics, semiconductor equipment, and more. The market they're going after is the large set of tasks where manipulation is too variable for traditional automation, labor is expensive or scarce, and a robot arm can drop into an existing workstation without requiring a factory redesign. As cobot hardware costs continue to fall, the economics get better every year.

We think Generalist is at the beginning of something important. The technology is improving on a curve we recognize, the team is world-class, and their deployments are already generating the data that makes the next generation of models better.