Genie Envisioner (AgiBot)

AgiBot's platform that unifies a video world model, a policy, a neural simulator, and evaluation into one system.

Genie Envisioner is a robot-manipulation platform released in August 2025 by the AgiBot team. Its core, GE-Base, is a diffusion model that generates video from a language instruction, trained on about 1 million clips (2,967 hours total) of dual-arm real-robot data from AgiBot World, learning to predict how the next frames will unfold. GE-Act uses a roughly 160-million-parameter flow-matching decoder to convert GE-Base's latent representation into actions; GE-Sim takes in an action and generates future frames, serving as a neural simulator for closed-loop evaluation; and EWMBench is used to measure the quality of this kind of world model. Code and weights are open-source. The September 2026 GE-Act 2.0 switched to training entirely from scratch on manipulation data, chaining together an autoencoder, a single-step visual planner, and an inverse dynamics model.

ExampleIn the paper, adapting to a new task with just about 1 hour of data lets GE-Act transfer to other bimanual platforms, such as AgileX's Cobot Magic, to carry out manipulation tasks.

Also called

GE-Base, GE-Act, GE-Sim, GE-Act 2.0, A Unified World Foundation Platform for Robotic Manipulation

Related

AgiBot · World Action Model · AgiBot World · EWMBench · Neural Simulator · AgiBot GO-1

Sources

Genie Envisioner (arXiv:2508.05635)
AgibotTech/Genie-Envisioner (GitHub)
GE-Act 2.0 项目主页 (Chinese)

As of

2026-09