Generalist’s $400M Bet on Robots That Can Actually Touch the World

Physical AI is moving fast.

But Andrew Barry says the next major breakthrough in robotics may come down to something deceptively hard: dexterity.

In this episode of Automated, Brian Heater speaks with Andrew Barry, co-founder and CTO of Generalist, about how the company is building general intelligence for the physical world.

They discuss why Generalist is starting with dexterous robots, why touching and manipulating objects is still one of the hardest problems in robotics, and why solving dexterity could unlock a much wider range of real-world automation.

Andrew also explains how Generalist built its own data capture devices, why real-world manipulation data matters so much, and how the company is training models on tasks people are already paying for today.

The conversation gets into one of the most surprising parts of modern robot learning: robots can sometimes complete tasks in ways they were never directly taught. Andrew describes why those moments changed how the team thought about robot learning, improvisation, and what these models may be capable of.

Brian and Andrew also cover GEN-1, the commercial reality behind robot demos, the challenge of flexible objects like cables, data flywheels in robotics, and what Andrew learned from working at Boston Dynamics and the Broad Institute.

If you want to understand why Generalist is getting so much attention in physical AI, this is the conversation.

KEY MOMENTS
(00:00) Why dexterity is the starting point for Generalist
(02:17) Why Generalist does not call its models VLAs or world models
(04:19) Why training from scratch matters
(05:56) How Generalist collects real-world robot data
(06:22) Why data gloves are different from teleoperation
(08:12) How the company scaled data collection
(12:02) Are workers training future automation systems?
(12:46) Why dexterity is the most valuable robotics problem
(14:35) Why physical AI may be following the GPT progression
(16:33) Why robot demos are not the same as commercial viability
(17:48) What changed from hard-coded robot behavior to model-based intelligence
(18:05) What Andrew learned from working on Spot at Boston Dynamics
(18:58) Why improvisational intelligence matters
(20:19) Why unrelated data can still improve robot performance
(20:45) How Generalist models handle new environments
(22:25) Why cable threading is so hard to automate
(23:29) What cross-embodiment means for robotics
(24:02) Why Generalist started when it did
(25:18) How the founders pitched a deep-tech robotics company
(26:00) Why the company focused its risk on dexterity
(27:24) Why Generalist is not starting by building humanoids
(29:07) Why the first year was all about data collection
(30:01) How video and language may fit into robot learning
(31:05) What the robot data flywheel may actually look like
(33:33) The robot did something they never taught it
(35:50) Why robots can learn human mistakes
(37:42) Andrew’s path from Boston Dynamics to the Broad Institute
(41:57) How machine learning was used in molecular biology
(43:12) Why transformers connect biology, language, and robotics

Connect with Andrew Barry
https://www.linkedin.com/in/andy-barry

Learn more about Generalist
https://generalistai.com/

We’d love to hear from you.
Have thoughts or guest suggestions?
Reach us at podcast@automate.org

You can find the transcript and more episodes of Automated at automated.fm

Unlock full access to Automated and explore everything automation.
Subscribe today and leave a review on YouTube, Apple Podcasts, and Spotify.
https://www.youtube.com/@automatedpodcast
https://podcasts.apple.com/us/podcast/automated-with-brian-heater/id1837762221
https://open.spotify.com/show/60olq6brlBEIJWggx2fMR6

You can also find us on:
LinkedIn https://www.linkedin.com/showcase/automated-podcast-by-a3/
Instagram https://www.instagram.com/automatedpod/

Subscribe to the Automated Newsletter:
https://www.automate.org/automation/automated-newsletter