GO-1 (Genie Operator-1)

GO-1 (Genie Operator-1) is a universal, generalist embodied foundation AI model developed by the Chinese company AgiBot, officially unveiled on March 10, 2025 in Shanghai. It redefines how robots see, understand and act in the real world — moving from rigid, task-specific automation toward flexible, generalist robotics. GO-1 is the heart of AgiBot's humanoid control stack, including the industrial G2 (unveiled October 16, 2025).

ViLLA (Vision-Language-Latent-Action) architecture

GO-1's novel architecture combines a Vision-Language Model (VLM) with a Mixture of Experts (MoE), forming the ViLLA framework. During inference the components work synergistically: the VLM first analyzes the scene and objects, the Latent Planner predicts k latent action tokens, and the Action Expert conditions a denoising process to produce the final control signals.

VLM (Vision-Language Model)

Uses internet-scale heterogeneous data as a solid foundation for scene and object understanding. Lets the robot interpret language instructions and ground them in the visual perception of its environment.

Latent Planner (MoE)

The first MoE expert — learns from cross-embodiment data and from human demonstrations, building general action understanding independent of any specific robot body. Its output is k latent action tokens that constitute an abstract action plan.

Action Expert (MoE)

The second MoE expert — trained on more than 1 million real-robot demonstrations (AgiBot World Colosseo). Performs high-frequency dexterous manipulation by converting the planner's latent tokens into concrete control signals via a denoising process.

Distinctive capabilities

GO-1 offers four unique properties: (1) Learning from Human Videos — learning from human video footage without teleoperation, (2) Few-shot Generalization — adapting to new tasks from minimal examples, (3) Cross-Embodiment Adaptation — transferring policies across robot platforms with different kinematics, (4) Continuous Self-Evolution — ongoing model improvement based on new operational data.

Applications and deployments

GO-1 drives the shift from rigid, task-specific automation toward universal generalist robotics. Deployed in manufacturing (precision assembly, automotive part transfer), logistics (parcel sorting), services (guided tours), and home automation. In the AGIBOT G2 humanoid it runs locally on the NVIDIA Jetson Thor T5000 (2,070 TFLOPS FP4) with control latency under 10 ms, paired with the GE-1 world model.