Zhiyuan Robot has officially declared the full open-sourcing of its versatile embodied foundation large model, GO-1. This groundbreaking model leverages the world's first-of-its-kind Vision-Language-Latent-Action (ViLLA) architecture. By integrating an intermediate semantic layer known as "Implicit Action Tokens," it effectively bridges the divide between image-text inputs and robotic action execution. This innovation significantly boosts the model's ability to generalize across diverse scenarios.

GO-1 has undergone extensive training on the AgiBot World dataset, which encompasses millions of trajectories from real-world robots, spanning 217 task categories and 106 different scenarios. In tests conducted on the Genie Sim and Libero simulation platforms, GO-1 has showcased exceptional performance. Moreover, it has been successfully adapted to function seamlessly with various hardware configurations, including the Songling Robot and Franka robotic arm.

Accompanying the launch of GO-1 is the Genie Studio development platform, which offers comprehensive, end-to-end assistance. This spans the entire process, from data acquisition and model training to simulation-based evaluation and real-world robot deployment. The platform facilitates the daily collection of up to a thousand data entries per machine, while also enhancing deployment efficiency by a factor of 2-3 compared to conventional methods.

Presently, GO-1 is available as an open-source resource on both GitHub and Huggingface platforms, granting developers unrestricted access at no cost.