Genie Operator-1 (GO-1)

GO-1 is the flagship model of Shanghai-based Zhiyuan Robotics, announced in spring 2025 and open-sourced that September. Its pitch is an architecture called ViLLA — Vision-Language-Latent-Action — which inserts a layer of implicit 'latent action' tokens between perception and control. Three pieces cooperate at inference: a VLM (InternVL-2B) that reads multi-view images, force signals, and language; a Latent Planner that turns those into latent action tokens (a 'chain of planning'); and a diffusion-based Action Expert that converts them into continuous high-frequency actions. The training fuel is AgiBot World, the company's own million-trajectory real-robot dataset across 217 tasks, plus human videos for the planner. Zhiyuan also ships Genie Studio, a dev platform for data collection, training, and sim evaluation. Capabilities like few-shot generalization and cross-embodiment transfer are company-reported; treat them as such until the open weights get independent mileage.

• Company-reported: few-shot generalization, cross-embodiment transfer, learning from human videos
• No independent benchmarks on the released weights found