FluxBisim

FluxBisim is a dual-arm collaborative manipulation simulation platform built on Isaac Sim 4.5.0. Integrated with the FluxVLA, it provides a complete simulation workflow for embodied AI tasks, covering scene selection, skill composition, asset configuration, data collection, model training, deployment, and multidimensional evaluation.

fluxbisim.mp4

🌟 Features

• Centers on task composition, flexibly combining primitive skills such as pick-and-place, opening and closing, rotation, pushing and pulling, and handover to create complex dual-arm collaboration tasks that reflect real-world application needs.
• Uses atomic controllers and the simulation engine to automatically generate high-quality training data at scale.
• Integrates the FluxVLA to support training, deployment, and evaluation of mainstream models in high-fidelity simulation environments.

Framework

📢 Latest News

[2026/05/28] 🔥 FluxBisim has been open-sourced.

🛠️ Installation

1. Clone the Repository

cd FluxVLA
git clone https://github.com/FluxVLA/FluxBisim.git

2. Download 3D Assets

The 3D assets are hosted in the Hugging Face Dataset repository limxdynamics/FluxBisimAssets. Download them to the assets directory in the project root:

cd FluxBisim
mkdir -p assets
hf download limxdynamics/FluxBisimAssets \
  --repo-type dataset \
  --local-dir assets

3. Install Isaac Sim

Download and extract Isaac Sim 4.5.0:

wget https://download.isaacsim.omniverse.nvidia.com/isaac-sim-standalone-4.5.0-linux-x86_64.zip
unzip isaac-sim-standalone-4.5.0-linux-x86_64.zip -d /path/to/install_dir/

We recommend configuring an alias for the Isaac Sim Python environment:

echo 'alias benchmark_python="/path/to/install_dir/isaac-sim-4.5.0/python.sh"' >> ~/.bashrc
source ~/.bashrc

4. Install ROS Noetic

sudo apt update
sudo apt install ros-noetic-desktop-full ros-noetic-cv-bridge \
  ros-noetic-sensor-msgs ros-noetic-std-msgs python3-rosdep
echo "source /opt/ros/noetic/setup.bash" >> ~/.bashrc
source ~/.bashrc

After installing ROS for the first time, initialize rosdep:

sudo rosdep init
rosdep update

5. Install Python Dependencies

Install the dependencies into the Python environment bundled with Isaac Sim:

benchmark_python -m pip install --upgrade pip
benchmark_python -m pip install rootutils PyYAML h5py scipy

Data Collection

cd /path/to/FluxVLA/FluxBisim
benchmark_python <collect_script> --env <env_name> --config <config_name> [--data_path <output_dir>]

Available tasks:

Task	Script	Environment	Config
Pick Fruit to Plat	data_collect/pick_place_fruit/fruit_pick_place_collect.py	kitchen, apartment	apple_pick_place_config.yaml, banana_pick_place_config.yaml, carrot_pick_place_config.yaml, cucumber_pick_place_config.yaml, mangosteen_pick_place_config.yaml, whiteradish_pick_place_config.yaml
Place Nut & Close Box	data_collect/close_box/box_close_collect.py	industry	box_close_config.yaml
Store Apple in Drawer	data_collect/pull_push_drawer/drawer_pull_push_collect.py	apartment	drawer_pull_push_config.yaml
Screw Pitcher Lid	data_collect/screw_pitcher_lid/pitcher_lid_screw_collect.py	apartmentshort	pitcher_lid_screw_config.yaml
Handover book	data_collect/handover_book/book_handover_collect.py	apartment, industry	book_handover_config.yaml

Example:

benchmark_python data_collect/pick_place_fruit/fruit_pick_place_collect.py \
  --env kitchen \
  --config banana_pick_place_config.yaml

Data collected automatically by FluxBisim is saved in HDF5 format. To use it for FluxVLA training, convert it to the LeRobot Dataset v2.1 format by following the data conversion guide.

📊 Benchmark

Leaderboard

Codebase	Pick Banana to Plate	Place Nut & Close Box	Store Apple in Drawer	Screw Pitcher Lid	Handover Book
FluxVLA (GR00T N1.5)	0.96	0.88	0.88	0.33	0.45
FluxVLA (Pi 0.5)	0.96	0.87	0.80	0.31	0.82

Dataset

The training dataset is hosted in the Hugging Face Dataset repository limxdynamics/FluxBisimData. It contains data automatically collected by FluxBisim and converted to the LeRobot Dataset v2.1 format, ready for FluxVLA training.

hf download limxdynamics/FluxBisimData \
  --repo-type dataset \
  --local-dir /path/to/FluxVLA/data/FluxBisimData

After downloading the dataset, set the data path in the FluxVLA training configuration to this directory.

Model Evaluation

Run the simulation-based evaluation environment in FluxBisim:

cd /path/to/FluxVLA/FluxBisim
PYTHON_BIN=/path/to/install_dir/isaac-sim-4.5.0/python.sh ./fluxbisim_benchmark.sh pick_place_banana

Supported evaluation tasks are close_box, handover_book, pick_place_banana, pull_push_drawer, and screw_pitcher_lid.

In a second terminal, launch FluxVLA inference:

cd /path/to/FluxVLA
./scripts/fluxbisim_eval.sh \
  configs/pi05/fluxbisim/pi05_paligemma_pick_place_banana_full_finetune.py \
  /path/to/checkpoint.pth

The task ID is inferred from the configuration file name by scripts/fluxbisim_eval.sh, so make sure the configuration matches the evaluation task.

🙏 Acknowledgements

This project builds on the contributions of the following open-source projects and communities: YCB Benchmark, Lightwheel SimReady, X-Humanoid ArtVIP, and Nvidia Omniverse USD. If we missed your project or contribution, please open an issue or pull request so we can properly acknowledge it.

This project also acknowledges AgileX Robotics for providing the URDF model of the dual-arm robotic platform used in this work.

Roadmap

• Release more evaluation metrics across three dimensions: spatial interaction, trajectory quality, and task outcomes.
• Support more dual-arm robots, such as Tron2.
• Strengthen domain randomization to improve data diversity and make model evaluation more comprehensive.
• Support data collection, model training, and evaluation for more scenes and tasks.
• Support data collection, model training, and evaluation for whole-body VLA models.