RoboNaldo Deploy

🌎English | 🇨🇳中文

​RoboNaldo Deploy​​ is a multi-policy robot deployment framework based on a state-switching mechanism. Currently, the included policies are designed for the ​​Unitree G1 robot (29-DoF)​​.

🤖 This is the deployment repository for RoboNaldo. The policy training code lives in the companion repository opendrivelab/RoboNaldo; this repository runs the trained policies on the real and simulated Unitree G1.

Simulation	Real robot
free_kick_sim.mp4	free_kick_real_small.mp4

Free-kick demo — RoboNaldo policy on a Unitree G1

Preface

• ​This deployment framework is only applicable to G1 robots with a 3-DOF waist. If a waist fixing bracket is installed, it must be unlocked according to the official tutorial before this framework can be used normally.​​

• It is recommended to remove the hands, as dance movements may cause interference.​

• When deploying real robots, if something goes wrong, it's probably the policy's fault—not your hardware. Don't waste time second-guessing your robot's physical setup.

• video instruction

Installation and Configuration

1. Create a Virtual Environment

It is recommended to run training or deployment programs in a virtual environment. We suggest using Conda to create one.

1.1 Create a New Environment

Use the following command to create a virtual environment:

conda create -n robomimic python=3.8

1.2 Activate the Virtual Environment

conda activate robomimic

---

2. Install Dependencies

2.1 Install PyTorch

PyTorch is a neural network computation framework used for model training and inference. Install it with the following command:

conda install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 pytorch-cuda=12.1 -c pytorch -c nvidia

2.2 Install RoboNaldo_Deploy

2.2.1 Download

Clone the repository via git:

git clone https://github.com/OpenDriveLab/RoboNaldo_Deploy.git

2.2.2 Install Components

Navigate to the directory and install:

cd RoboNaldo_Deploy
pip install -r requirements.txt

2.2.3 Install unitree_sdk2_python

git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
cd unitree_sdk2_python
pip install -e .

Documentation Map

For real-robot onboard setup, start with onboard/docs/README.md. It explains the runtime architecture, topic ownership, and the recommended reading path for new contributors.

• Environment setup: onboard/docs/INSTALL_AGENT_GUIDE.md
• Real-robot bring-up topology: tools/start_tmux_layout.sh
• Camera module guide: onboard/perception/camera/README.md
• LiDAR module guide: onboard/perception/lidar/README.md
• Perception architecture: onboard/docs/PERCEPTION_ARCHITECTURE.md

---

Running the Code

1. Run Mujoco Simulation

Standard scene:

python deploy_mujoco/deploy_mujoco.py

FreeKick scene (includes ball, required for FreeKick policy):

python deploy_mujoco/deploy_mujoco.py --config-name mujoco_freekick

2. Policy Descriptions

Mode Name	Trigger Keys	Description
PassiveMode	F1 (real) / L2 release (sim)	Damping protection mode
FixedPose	Start	Position control reset to default joint values
LocoMode	B	Stable walking control mode
AMP	A	AMP locomotion policy (enters run mode by default)
FreeKick	R1	Ball-kicking/scoring policy (547-dim obs, 5-frame history, requires onboard LiDAR)
BeyondMimicMJ	D-pad DOWN	Stand-to-lying imitation policy (MuJoCo-trained, with reference motion tracking)
StandUpMJ	D-pad UP	Lying-to-standing imitation policy (MuJoCo-trained, with reference motion tracking)
Pinocchio1.6MJ	R2	MuJoCo imitation policy backed by g1_result_pinocchio_1_6_mj.yaml
BeyondMimic	—	Still kept in the repo, but not bound to a default controller shortcut

Manual Bias Adjustments

Both sim and real support runtime Y-axis bias to correct perception offsets without restarting:

Action	Keys	Range	Notes
Target Y +5 cm (left)	Hold L1 + tap D-pad Left	±1.5 m	Sim + real
Target Y −5 cm (right)	Hold L1 + tap D-pad Right	±1.5 m	Sim + real
Ball Y +5 cm (left)	Hold L2 + tap D-pad Left	±1.5 m	Real only
Ball Y −5 cm (right)	Hold L2 + tap D-pad Right	±1.5 m	Real only

Bias is in the pelvis body frame (+Y = robot's left). It shifts the position fed to the policy without affecting the raw sensor reading. Each change prints to the terminal and is always visible in the Sensor Dashboard orange "Current Bias" card.

---

3. Operation Instructions in Simulation

1. Connect an Xbox controller.
2. Run the simulation program:

python deploy_mujoco/deploy_mujoco.py

3. Press the Start button to enter position control mode.
4. Press B to enter LocoMode, then press BACKSPACE in the simulation to make the robot stand. Afterward, use the joystick to control walking.
5. The simplified single-button policy map is:
   • A -> AMP (run mode)
   • B -> LocoMode
   • D-pad DOWN -> BeyondMimicMJ
   • D-pad UP -> StandUpMJ
   • R2 -> Pinocchio1.6MJ
   • R1 -> FreeKick
6. Use --config-name mujoco_freekick when entering FreeKick so the simulation loads the ball scene.
7. BeyondMimicMJ / StandUpMJ / Pinocchio1.6MJ can switch directly between each other with the same single-button bindings.
8. Release L2 at any time for damping protection (PassiveMode), or press Start to return to FixedPose.
9. BeyondMimic is still present in the repository, but it is intentionally left without a default controller shortcut. Unused legacy policies such as Dance / Kick / KungFu / ASAP / HOST have been deleted from the repository.

---

4. Real Robot Operation Instructions

1. Power on the robot and suspend it (e.g., with a harness). First hold L2+B to enter damping protection mode (the head light changes from purple to orange), then hold L2+R2 to enter debug mode (the head light changes from orange to yellow).

2. Recommended (onboard Orin): Use the tmux launcher, which starts the C++ bridge, policy runtime, perception services and sensor dashboard in one shot:

   bash tools/start_tmux_layout.sh

   Then in the deploy_policy pane, press Start to enter position control mode.

   Alternatively, run the C++ bridge and Python policy manually:

   BRIDGE_NETWORK_INTERFACE=eth0 bridge/build/cpp_bridge_main   # terminal 1
   python bridge/python/deploy_policy.py                          # terminal 2

   For quick testing without the C++ bridge (Python bridge prototype):

   python bridge/python/deploy_bridge_py.py   # terminal 1
   python bridge/python/deploy_policy.py       # terminal 2

   Simple mode (no Orin / direct Ethernet): python deploy_real/deploy_real.py

3. The same single-button mapping applies on the real robot: A / B / D-pad DOWN / D-pad UP / R2 / R1. Press F1 at any time for damping protection (PassiveMode).

4. FreeKick policy (R1) — additional real-robot steps: The FreeKick policy depends on onboard ball perception. Start the perception services before activating FreeKick; raw LiDAR/camera observations are fused by onboard/perception/ball_fuser.py, which publishes the final DDS topic rt/ball_state. Do not activate the FreeKick policy without the perception services running.

5. Sensor Dashboard: After starting tools/start_tmux_layout.sh, open http://<robot-ip>:8091/ in a browser. The dashboard shows all sensor positions (target, cam/lidar/fused ball, corrected positions) and the active bias values in the orange "Current Bias" card. To start it manually:

   python onboard/perception/debug/sensor_dashboard.py

   See bridge/README.md and bridge/VALIDATION.md for full bridge setup details.

---

Important Notes

1. Framework Compatibility Notice

The current framework does not natively support deployment on G1 robots equipped with Orin NX platforms. Preliminary analysis suggests compatibility issues with the unitree_python_sdk on Orin systems. For onboard Orin deployment, we recommend the following alternative solution:

• Replace with unitree_sdk2 (official C++ SDK)
• Implement a dual-node ROS architecture:
  • C++ Node: Handles data transmission between robot and controller
  • Python Node: Dedicated to policy inference
• This repository now includes a minimal bridge prototype under bridge/ that can serve as a reference implementation for that split.

2. Mimic Policy Reliability Warning

The Mimic policy does not guarantee 100% success rate, particularly on slippery/sandy surfaces. In case of robot instability:

• Press F1 to activate PassiveMode (damping protection)
• Press Select to immediately terminate the control program

3. Notes For The Retained Policies

The default deploy entry points are now limited to LocoMode / AMP / FreeKick / BeyondMimicMJ / StandUpMJ / Pinocchio1.6MJ.

⚠️ Important Precautions:

• Palm Removal Recommended: Some original training setups did not account for palm collisions (the author's G1 had no palms)
• Prefer Safe Transition Paths: Before and after testing a retained policy, prefer returning to FixedPose or PassiveMode
• Use Human Supervision: For first-time real-robot validation, keep the robot suspended and maintain manual supervision throughout

4. Other Movement Advisories

Only the retained deploy policies remain in the repository. Other historical policy implementations were removed to keep the deployment surface minimal.

5. Strong Recommendation

Always master operations in simulation before attempting physical robot deployment.

Acknowledgements

This deployment framework is built on top of ccrpRepo/RoboMimic_Deploy. We thank the authors for open-sourcing the original state-switching multi-policy deployment framework that this repository extends.

It is the deployment counterpart of RoboNaldo — training code: opendrivelab/RoboNaldo.