GE-1

GE-1 is a video-generative world model developed by the Chinese company AgiBot, released in August 2025. It is designed as a partner to the GO-1 foundation model in AgiBot's humanoid control stack — in the industrial G2 (launched October 16, 2025) GE-1 is responsible for predicting future scenarios in time and space, allowing the robot to rehearse actions in a virtual environment before executing them in the real world.

Closed-loop architecture

GE-1 combines three components in a single closed loop: (1) video generation — predicting future observation frames conditioned on robot actions, (2) policy learning — using simulated future scenarios to tune the control policy, (3) simulation evaluation — validating planned actions in the virtual world before physical execution. Together it realizes full end-to-end reasoning from seeing, through thinking, to acting.

Pairing with GO-1

GE-1 does not replace GO-1; it complements it. GO-1 (ViLLA: VLM + Latent Planner + Action Expert) emits control signals for the current action, while GE-1 provides the prediction horizon as generated video and simulation. This two-model setup is the heart of the AI in the G2 humanoid — running locally on the NVIDIA Jetson Thor T5000 (2,070 TFLOPS FP4) with total control latency below 10 ms.

Position in the field

GE-1 fits in the broader wave of world models for robotics (world models, action-conditioned video generation), where generated predictions replace costly or unsafe physical trials. Similar approaches: NVIDIA Cosmos, Google Genie 3, World Action Model. GE-1 stands out by being integrated into a ready production stack (GO-1 + G2) and by its claimed industrial maturity — the model is not just a research prototype.