AgiBot GO-1

At a glance

Overview

GO-1 is AgiBot's generalist embodied foundation model, notable for pioneering the ViLLA (Vision-Language-Latent-Action) architecture — an evolution of standard VLA. It combines a Vision-Language Model, a Latent Planner trained on cross-embodiment and human-operation data, and an Action Expert (MoE), with latent action tokens quantized via VQ-VAE. A lighter GO-1-Air variant is also published. Open weights are on Hugging Face.

What it's used for

Cross-embodiment manipulation research; teams adapting a generalist base to their own robot arms or humanoid platforms.

Primary sources

• Hugging Face: https://huggingface.co/agibot-world/GO-1
• GitHub: https://github.com/OpenDriveLab/AgiBot-World
• Official site: https://www.agibot.com/