AgiBot Genie Envisioner

Architecture

Action-driven world model with policy learning, evaluation, and simulation

Key facts

Unified world platform components (primary arXiv)

The Genie Envisioner technical report (arXiv:2508.05635) describes GE as a unified world foundation platform for robotic manipulation integrating policy learning, evaluation, and simulation in one video-generative framework. Components named include GE-Base (instruction-conditioned video diffusion), GE-Act (flow-matching action decoder), GE-Sim (action-conditioned neural simulator), and EWMBench. Treat architecture claims as author-stated ([arXiv](https://arxiv.org/pdf/2508.05635))

Open-source launch coverage (primary Global Times / AgiBot)

Global Times coverage of AgiBot's August 2025 launch says Genie Envisioner integrates prediction, policy learning, and neural simulation in a video-generative framework, with GE-Base trained on about 3,000 hours of video-language data spanning more than 1 million real-world robotic manipulation episodes, and that AgiBot planned to open-source code, models, and benchmarks. Treat training-scale figures as company-stated via the report ([Global Times](https://www.globaltimes.cn/page/202508/1340873.shtml))

Unified world foundation platform (primary arXiv)

The Aug 2025 arXiv abstract for Genie Envisioner (GE) from the AgiBot Genie Team describes a unified world foundation platform for robotic manipulation integrating policy learning, evaluation, and simulation in one video-generative framework, with GE-Base as an instruction-conditioned video diffusion model, GE-Act mapping latents to action trajectories, GE-Sim as an action-conditioned neural simulator, and EWMBench for evaluation. Authors state code, models, and benchmarks will be released publicly ([arXiv](https://arxiv.org/abs/2508.05635))

Open-source repository (primary GitHub)

The AgibotTech/Genie-Envisioner GitHub repository hosts the open Genie Envisioner codebase associated with the paper and project site. Treat repository contents as primary engineering release ([GitHub AgibotTech](https://github.com/AgibotTech/Genie-Envisioner))

Unified world foundation platform for manipulation (project site)

AgiBot's Genie Envisioner project materials describe a unified world foundation platform for robotic manipulation that integrates policy learning, simulation, and evaluation in a video-generative framework ([Genie Envisioner](https://genie-envisioner.github.io/)).

GE-Base / GE-Act / GE-Sim technical report (arXiv Aug 2025)

The August 2025 technical report frames Genie Envisioner components including GE-Base video diffusion, GE-Act action policy inference, and GE-Sim closed-loop simulation, with EWMBench evaluation. Treat benchmark scores as author-stated ([arXiv](https://arxiv.org/pdf/2508.05635)).

Developed by (1)

• AgiBot

Common questions

Sources (5)

1. https://arxiv.org/pdf/2508.05635 ↗
2. https://www.globaltimes.cn/page/202508/1340873.shtml ↗
3. https://arxiv.org/abs/2508.05635 ↗
4. https://github.com/AgibotTech/Genie-Envisioner ↗
5. https://genie-envisioner.github.io/ ↗

Related brains

• UnifoLM5 models · open
• 1X World Model2 models
• DYNA-21 model
• PonyWorld (Pony.ai Virtual Driver)1 model
• FLUX 3 Actionopen source
• Genesis World Model