GAIllustrationNo official image on file. Showing a schematic category illustration.
Generalist's robot foundation model, claimed to master simple physical tasks with very little task-specific robot data.
Key specs
| Claims | 99% success rates, 3x faster, 1 hour of robot data per task |
|---|
Technical details
GEN-1 is Generalist AI's embodied foundation model, scaled from GEN-0 and trained from scratch on more than 500,000 hours of physical-interaction data captured with wearable devices on humans, with no robot data in pretraining. Generalist defines 'mastery' as reliability, speed and improvisation, and reports 99% average success on tasks where GEN-0 reaches 64%, using roughly one hour of robot data per task. It claims ~3x faster task completion than prior SOTA (e.g. 12.1 s box folds) via learning from experience and a new inference approach. Access is via early-access partnerships.
| Architecture | Large multimodal model that emits actions in real time; described as a system combining pre-training advances, post-training, RL from experience, multimodal human guidance and new inference-time techniques ('Harmonic Reasoning', paged attention for real-time inference). |
|---|
| Inputs outputs | Multimodal robot observations in; real-time robot actions out. |
|---|
| Training data | Trained from scratch on over 500,000 hours of real-world physical-interaction data from low-cost wearable devices on humans; base pretraining contains no robot data; ~1 hour of robot data per task for adaptation. |
|---|
| License weights | Proprietary; early-access partners. |
|---|
| Benchmarks claimed | | Benchmark | Average success on reported tasks (vs GEN-0 fine-tuned) |
|---|
| Result | 99% (GEN-0: 64%; from scratch: 19%) |
|---|
| Source url | generalistai.com/…/gen-1 |
|---|
|
|---|
| Sources | |
|---|
Benchmarks, where listed, are as claimed by the vendor.
