---
title: FLUX 3 Action SO-101 Sim
emoji: 🦾
colorFrom: yellow
colorTo: gray
sdk: gradio
sdk_version: 6.28.0
app_file: app.py
python_version: "3.12"
short_description: FLUX 3 Action drives a simulated SO-101 arm in MuJoCo
startup_duration_timeout: 1h
models:
  - black-forest-labs/flux-3-action-so101
  - black-forest-labs/flux-3-action-base
---

Closed-loop demo of [FLUX 3 Action SO-101](https://huggingface.co/black-forest-labs/flux-3-action-so101) controlling a simulated SO-101 arm in MuJoCo.

Robot model and meshes: [TheRobotStudio/SO-ARM100](https://github.com/TheRobotStudio/SO-ARM100) (`so101_new_calib_camera.xml`, Apache-2.0).

How it works:
- Robot: `so101_new_calib_camera.xml` with decimated meshes, joints driven in LeRobot degrees (new calibration), wrist camera upright like the standard SO-101 mount.
- Normalization: the released checkpoint ships one dataset's quantiles (an SO-100 kitchen set in the old calibration). This demo instead uses the median quantiles of the new-calibration SO-10x datasets in the training corpus, which predicts real SO-101 episodes about 2x more accurately offline.
- Speed: the whole closed loop runs in one ZeroGPU call; DiT blocks load AoTI kernels from `multimodalart/flux-3-action-so101-aoti` (set `FLUX3_AOTI=0` to force eager).