Skip to content

GO-1

Carry vision and language inputs over to robot actions

Visit official site

What GO-1 is

The model is described as a general embodied foundation model from AgiBot that uses the ViLLA architecture with a VLM and an MoE system to turn vision and language inputs into robot actions.

What you can do with it

  • Predict latent action tokens to link inputs with action execution
  • Generalise to new tasks with very little data or zero samples
  • Learn from human video to understand human behaviour
  • Move across robot forms and refine performance through a data feedback loop

Application scenarios

  • Service robots in commercial settings handle cleaning and carrying
  • Robot operation flows are optimised in industrial settings
  • Household robots pour water and clear a table

Getting started

  • Obtain the model and documents from the AgiBot site or the support email
  • Pick a robot body that fits the task
  • Fine-tune with the AgiBot World dataset or custom data
  • Deploy the model to the robot system with hardware and software compatibility

Pros & cons

✓ What we like

  • Human video and real robot data are both used for learning
  • The model adapts across different robot bodies

! What to watch out for

  • Action output depends on the ViLLA architecture and latent action tokens
  • Adaptation starts from the AgiBot World dataset or custom data

FAQ

What is ViLLA?

Vision-Language-Latent-Action is named as the architecture that predicts latent action tokens to bridge input and action execution.

How are new tasks handled?

The material lists few-shot generalisation, including very little data or zero samples.

Last reviewed: 2026-11-15

More AI chatbot tools

View all →

How we review