Tired of projects that release only models but no deployment code? RoboJuDo provides a unified framework that makes policy deployment straightforward and practical.
Our framework highlights:
-
Out-of-the-box: After setting up RoboJudo, multiple policies can be deployed on both simulation and real robots in minutes: Quick Start.
-
Decoupled & Modular Design: With a Python-first design, RoboJuDo makes fast prototyping easy. Environment, Controller, and Policy are modular and freely composable, while minimal code changes allow seamless adaptation across robots and setups: See how we achieve this: Add a new module.
-
Multi-policy switching: Seamlessly switch between different policies during a task. Try this: Multi-Policy Switching.
-
Light-Weight: Our framework is lightweight, after 5 minutes of setup, it runs smoothly onboard. By UnitreeCpp, RoboJuDo runs on Unitree G1 without the need for an Ethernet cable.
📓Content
🗺️Roadmap
2026.03 Update: We added built-in support for deploying the ProtoMotions G1 tracker in RoboJuDo. Thanks to NVLabs for the great work on ProtoMotions, and thanks to Chen Tessler and Yifeng Jiang for contributing this integration.
📄Introduction
This repository provides a deployment framework for humanoid robots, supporting the use of different policies across different environments (real robots and simulation).
We decouple the controller, environment, and policy, making it easy for users to add their own policies or environments.
Experiment configurations can be organized through config files.
The main modules of RoboJuDo consist of:
- 🎮 Controller: A collection of control signals. It receives external inputs (e.g., joystick, keyboard, motion sequences) and forwards them as
ctrl_data to the pipeline.
- 🤖 Environment: The execution environment (e.g., Mujoco, real robot). It processes actions provided by the policy and sends real-time sensor data as
env_data to the pipeline.
- 🌐 Policy: A trained control policy (from various wbc & locomotion works). It generates actions based on information from both the environment and the controller.
Currently, RoboJuDo supports the following policy–environment combinations:
🖥️ means policy is ready for simulation, while 🤖 means policy has been tested on real robot.
🛠️Easy Setup
RoboJuDo supports multiple platforms, officially tested on Ubuntu, macOSand Windows.
Robot onboard PCs are also supported.
1️⃣ Basic Installation
Step 1: Clone the repository and create a Python environment
git clone https://github.com/HansZ8/RoboJuDo.git
cd RoboJuDo/
# Example using conda
conda create -n robojudo python=3.11 -y
conda activate robojudo
Step 2: Install RoboJuDo
# Optional, install cpu version for speed up
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -e .
2️⃣ Install Optional Modules
RoboJuDo is a plug-and-play framework. After a minimal default installation, you can selectively configure and install only the modules you need.
Step 0: [Optional] Install Robot SDK
You can skip this for sim2sim and development.
If you plan to control a real robot, install the corresponding SDK.
For example, see unitree_setup.md for Unitree robots.
Step 1: Configure modules
Edit submodule_cfg.yaml to select modules, by setting install as true.
As default, mujoco_viewer is selected for sim2sim.
Step 2: Install modules
# Install all required modules
python submodule_install.py
# Or specify modules to install with args
# python submodule_install.py unitree_cpp
📖Quick Start
RoboJuDo is a modular framework where tasks can be flexibly defined by composing configuration files.
In the following, we use the deployment on G1 as an example.
- Run Sim2Sim
- Run Sim2Real
- Deploy More Policies✨
Run RoboJuDo on Simulation
Begin your journey with unitree g1 sim2sim.
A Xbox controller is needed for control.
# run the default g1 sim2sim cfg
python scripts/run_pipeline.py
You can control the motivation using any Xbox controller:
left axes move forward/backward/lfet/rightright axes turn left/right
Run RoboJuDo on Real Robot 🤖
Alert & Disclaimer ⚠️⚠️⚠️
Before deployment, you’d better first purchase accident insurance to cover any potential incidents that may occur during real-world operation. Policies could cause ⚠️violent motions⚠️ when losing balance. Always verify that the emergency stop button (e.g., A for default config) works properly.
Unless you have strong sim-to-real expertise and rigorous safety measures, DO NOT run these models on real robots. They are provided for research only, and we disclaim any responsibility for harm, loss, or malfunction.
Robot Setup
Follow our setup guide to set up the robot sdk on your computer or robot.
Start RoboJuDo
Open g1_cfg.py and modify the g1_real config.
Edit the env_type and net_if according to your robot sdk setup.
class g1_real(g1):
env: G1RealEnvCfg = G1RealEnvCfg(
env_type="UnitreeEnv", # For unitree_sdk2py
# env_type="UnitreeCppEnv", # For unitree_cpp, check README for more details
unitree=G1UnitreeCfg(
net_if="eth0", # note: change to your network interface
),
)
Refer to official guide to prepare and start the robot.
Then start the pipeline on the real robot:
python scripts/run_pipeline.py -c g1_real
Your robot should move into default pos.
During the preparation, put your robot on the ground.
You can control the real robot using the Unitree controller:
A button: Emergency stop. The robot immediately switches to damping mode. Be careful.left axes move forward/backward/let/rightright axes turn left/right
Deploy More Policies
💡Now you’re familiar with RoboJuDo’s config design, it’s time to experience the amazing variety of policies!
BeyondMimic & ASAP
Try the out of box experience of BeyondMimic and ASAP:
python scripts/run_pipeline.py -c g1_beyondmimic
python scripts/run_pipeline.py -c g1_asap
check documentation BeyondmimicPolicy and AsapPolicy for more details.
Multi-Policy Switch
g1_switch config in g1_cfg.py is equipped with Multi-Policy Pipeline.
python scripts/run_pipeline.py -c g1_switch
Xbox Controller:
left axes move forward/backward/left/right
right axes(left/right) turn left/right
Switch between Unitree Policy and AMO Policy:
RB + Dpad[Down] switch to Unitree PolicyRB + Dpad[Up] switch to AMO Policy
Loco-Mimic Policy Switch with Interpolation
For deploying Motion Mimic Policies with Locomotion as backup, we built LocoMimicPipeline for multi-policy switching with interpolation,
Check g1_locomimic config in g1_cfg.py, and more fancy locomimic configs in g1_loco_mimic_cfg.py.
python scripts/run_pipeline.py -c g1_locomimic_beyondmimic
python scripts/run_pipeline.py -c g1_locomimic_asap
We have the same Keyboard control as ASAP:
[ to switch to MotionMimic] to switch to LocoMotion; toggle next mimic policy' toggle prev mimic policy
More Policies
We also provide config files for other policies, check config_g1 and config_h1 for more details.
ASAP
In RoboJuDo, we have fully replicated ASAP’s Sim2Real workflow, including all motions.
Please refer to g1_locomimic_asap_full in g1_loco_mimic_cfg.py. This highlights the modular advantages of our framework.
ProtoMotions Tracker
Please first follow the official workflow from NVLabs / ProtoMotions and clone the ProtoMotions repo as a sibling directory named protomotions.
Then you can run the built-in RoboJuDo config directly:
python scripts/run_tracker_pipeline.py -c g1_protomotions_tracker \
--motion-path assets/motions/g1/g1_bones_seed_mini.pt \
--motion-index 0
For more details, check ProtoMotionsTrackerPolicy.
🧩Develop and Contribute
Add a new module
Refer to the documentation on Policy, Controller, Env, create and deploy your own policy in minutes.
(By the way, deploying AMO takes only 30 minutes, and GMT about 1 hour in our framework.)
Or simply create an issue — we will include updates in future releases!
Contribute to our project
We warmly welcome contributions from the community. Let’s build a strong and open ecosystem for RoboJuDo together!
🔗Citation
If you find our work useful, please cite our GitHub repository:
@misc{li2026fromw1generalhumanoidwholebody,
title={FRoM-W1: Towards General Humanoid Whole-Body Control with Language Instructions},
author={Peng Li and Zihan Zhuang and Yangfan Gao and Yi Dong and Sixian Li and Changhao Jiang and Shihan Dou and Zhiheng Xi and Enyu Zhou and Jixuan Huang and Hui Li and Jingjing Gong and Xingjun Ma and Tao Gui and Zuxuan Wu and Qi Zhang and Xuanjing Huang and Yu-Gang Jiang and Xipeng Qiu},
year={2026},
eprint={2601.12799},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2601.12799},
}
@misc{RoboJuDo,
author = {Zihan Zhuang and Yi Dong and Peng Li},
title = {A plug-and-play deploy framework for robots. Just deploy, just do.},
url = {https://github.com/HansZ8/RoboJuDo},
year = {2025}
}
or star our repo😁
- Unitree SDK2 Python: used for implementing
UnitreeEnv.
- PHC: used for implementing the
MotionCtrl module for OmniH2O.
- UnitreeCpp: our pybind of
unitree_sdk2 used in UnitreeCppEnv.
- ZED Proxy: ZED Camera Odometry Service.
- ProtoMotions: GPU-accelerated simulation and learning framework by NVLabs.
评论