RoboCat
Gato-based self-improving robot agent that learns new tasks from 100-1,000 demos and retrains on its own practice data.
Multi-embodiment agent built on Gato that learns a new task or arm from 100-1,000 human demonstrations, generates its own extra data (about 10,000 practice attempts), and retrains. Success on unseen tasks rose from 36% to 74% across versions.
- Date
- Tuesday, 20 June 2023
- Lab
- Google DeepMind
- Kind
- paper
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Unseen-task success (500 demos/task) | 36% to 74% first to final version | company |
Sources
This record was checked against its sources on 6 October 2026. How we check