ABAB Pedia

Generalist AI

Overview

Generalist AI describes its work as embodied foundation models, initially focused on dexterity, and lists locations in the California Bay Area and Boston.⁠[1]

Contents8 sections
Key facts

GEN-0

The company introduced GEN-0, reporting pretraining on over 270,000 hours of real manipulation data and using Harmonic Reasoning to coordinate sensing and action.⁠[2]

GEN-1

The company introduced GEN-1, reporting training on more than 500,000 hours of real interaction data and task adaptation with about one hour of robot data. Its evaluations emphasized reliability, speed and improvisation.⁠[3]

Funding

8VC backed the $400 million round, listing CEO Pete Florence, chief scientist Andy Zeng and CTO Andrew Barry.⁠[4]

End Effectors

The company reported adapting GEN-1 to varied end effectors and analyzing weight changes after fine-tuning on different tools; its pretraining data covered about 9,000 hand variants.⁠[5]

GEN-1.5

The company introduced GEN-1.5, reporting in-context learning from short demonstrations without gradient updates. Across ten simple, short-horizon manipulation tests, average success was 59%; after ten training steps on five minutes of data per task, it was 83%.⁠[6]

Related Organizations, Websites, and Public Accounts

Sources

  1. About / Generalist (opens in a new window)
  2. GEN-0 / Embodied Foundation Models That Scale with Physical Interaction (opens in a new window)
  3. GEN-1: Scaling Embodied Foundation Models to Mastery (opens in a new window)
  4. Announcing our Investment in Generalist (opens in a new window)
  5. Towards Machines with a Thousand Hands (opens in a new window)
  6. GEN-1.5: Embodied Foundation Models are One-Shot Learners (opens in a new window)
  7. Generalist (@GeneralistAI) / X (opens in a new window)
  8. Generalist (opens in a new window)
  9. Generalist - YouTube (opens in a new window)