Genie Envisioner
AgiBot, China 路 August 2025
AgiBot (Genie Team), with LV-NUS Lab and Beihang University (BUAA)
World model 路 Open weights, RAIL-M 路 Also written GE, GE-Base, GE-Act, GE-Sim
- Parameters
- not published
- No total is published.
- GPU memory
- not published
- weights at bf16, computed
- Inference latency
- 200 ms
- an onboard NVIDIA RTX 4090, one 54-step action chunk
- Weights
- RAIL-M
- RAIL-M (LTX Video 0.9 licence) for the GE-Base checkpoints, which derive from ltx-video-2b-v0.9.
What it is
Genie Envisioner is AgiBot's world model platform for manipulation, built around one instruction-conditioned video diffusion model. GE-Base, initialised from LTX-Video 2B or Cosmos2 2B, learns from about 3,000 hours of AgiBot World dual-arm episodes seen through a head camera and two wrist cameras. GE-Act, a 160M flow matching decoder, turns its latent video features into 54-step action chunks in 200 ms on an onboard RTX 4090, and GE-Sim reuses the model as an action-conditioned neural simulator for closed-loop policy tests. The paper reports GE-Act ahead of GR00T N1, 蟺0 and UniVLA on its real tasks, but gives most real-world numbers only in figures. A second generation simulator, GE-Sim V2, appeared on Hugging Face in June 2026 with 2B checkpoints for G01, G02 and YAM setups.
Architecture
- Backbone
- Video diffusion transformer initialised from LTX-Video 2B (or Cosmos2 2B), with a frozen T5-XXL text encoder, predicting three synchronized views (head and two wrists).
- Action head
- GE-Act: a 160M parameter flow matching action decoder on the video model's latent features. 54 actions at 30 Hz per chunk; video runs at 5 Hz and actions at 30 Hz, with one flow matching step for video and five for actions.
- Action chunk
- 54 steps per forward pass
- Control rate
- 30 Hz
- Parameters
- No total is published. The video model starts from a 2B DiT (LTX-Video 2B or Cosmos2 2B) and GE-Act adds 160M parameters.
- Pretraining data
- AgiBot-World-Beta: about one million dual-arm episodes, roughly 2,967 hours, from a head camera and two wrist cameras.
- Embodiments
- AgiBot G1, Agilex Cobot Magic, Dual Franka
What hardware it needs
This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.
| Inference latency, reported | 200 ms on an onboard NVIDIA RTX 4090, one 54-step action chunk source |
| Control rate | 30 Hz |
No memory figure. Training GE-Base took 32 A100 GPUs for about seven days in stage I and about three days in stage II.
Reported results
Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.
Simulation
No results in this category are published for this model.
Real world
No results in this category are published for this model.
Fine tuned tasks
No results in this category are published for this model.
On real hardware
AgiBot G1 tasks (making a sandwich, pouring tea, cleaning a table, heating food in a microwave, packing detergent) are compared with UniVLA and GR00T N1, and cloth and box folding on Agilex Cobot Magic and Dual Franka with GR00T N1, 蟺0 and UniVLA from 250 demonstrations (about one hour) per task. The text states that GE-Act outperforms all three baselines and that UniVLA and GR00T N1 reach 0 percent on the complex Agilex tasks; per-task rates are only in figures.
Fine tuning it yourself
New embodiments are adapted from about one hour of teleoperated demonstrations (250 per task in the paper). GE-Base checkpoints in a fast and a slow variant are on Hugging Face, and the code is in the AgibotTech GitHub repository.
Where it helps, where it does not
Strengths
- Covers the whole loop in one model family: video prediction, action decoding and a neural simulator.
- Fast action decoding for a video based policy: a 54-step chunk in 200 ms on an RTX 4090.
- Pretrained on about 3,000 hours of real dual-arm data with three camera views.
- Comes with EWMBench, a benchmark for embodied world models covering visual fidelity, physical consistency and instruction-action alignment.
Limits
- Licences: code and data CC BY-NC-SA 4.0 (non-commercial), checkpoints under the LTX Video 0.9 RAIL-M licence with use restrictions.
- Most real-world results are given only as figures, so they cannot be quoted as numbers.
- Pretraining data comes from AgiBot hardware; other robots need their own demonstrations.
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://arxiv.org/abs/2508.05635
- https://arxiv.org/html/2508.05635v3
- https://huggingface.co/agibot-world/Genie-Envisioner-v1.0
- https://huggingface.co/agibot-world/Genie-Envisioner-Sim-v2.0
- https://github.com/AgibotTech/Genie-Envisioner-V1
- https://genie-envisioner.github.io/
- https://huggingface.co/Lightricks/LTX-Video/blob/main/ltx-video-2b-v0.9.license.txt
Entry last checked 2026-11-12.
Try a policy on a real arm
A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.