Phase 1: Whole Body Teleoperation

Running the Sonic Simulation on a PC

Minimum System Requirements

CPU: x86_64 architecture
GPU: Nvidia RTX 3070/4060, 8 GB VRAM
RAM: 16 GB
Disk Space: 200 GB

Install Dependencies

Install Docker Tutorial

tip

You need to add your user to the docker group.

sudo usermod -aG docker $USER # Reboot required after running this

Install NVIDIA Container Toolkit Tutorial

Download ioenv

git clone https://github.com/ioai-tech/ioenv_cli.git

Load environment variables

echo 'source <path_to_ioenv_cli>/ioenv_cli/ioenv.sh' >> ~/.bashrc # Replace <path_to_ioenv_cli> with the local path to ioenv_cli
source ~/.bashrc

List available images with ioenv list

ioenv list

ENVIRONMENT          IMAGE      CONTAINER    STATUS    
-----------          -----      ---------    ------    
isaaclab             yes        --           --        
mujocosim            --         --           --        
onboard              --         --           --        
sonic_onboard        --         --           --        
sonic_sim            --         --           --   
sonic_train          yes        yes          running   

Pull an image with ioenv pull <ENVIRONMENT>

ioenv pull sonic_sim

# Mirror download for users in regions with limited Docker Hub access
ioenv pull sonic_sim --from-mirror

Running

# First terminal
ioenv run sonic_sim

cd /opt/GR00T-WholeBodyControl
source .venv_teleop/bin/activate
python gear_sonic/scripts/run_sim_loop.py

# Second terminal
ioenv run sonic_sim

cd /opt/GR00T-WholeBodyControl/gear_sonic_deploy
./deploy.sh sim --input-type zmq_manager

## If using keyboard input instead
./deploy.sh sim

# Third terminal
ioenv run sonic_sim

cd /opt/GR00T-WholeBodyControl
source .venv_teleop/bin/activate
python gear_sonic/scripts/pico_manager_thread_server.py --manager

Running XRoboToolkit on the Pico

Installation

Package URL: XRoboToolkit-PICO-1.1.1.apk
Download the package to your laptop, then connect the Pico to the laptop via USB.
Drag the package into the Pico's Download folder.
Disconnect the USB cable and install from the Pico home screen.

Calibration

Open the motion tracker.
Select the 3-tracker wearing mode.
Click "Start Pairing", then press and hold the button on the Pico motion tracker for 6 seconds to enter pairing mode.
Click the button to calibrate.

Settings

Change PC Service to the IP of the device running SONIC.

Deploying Sonic on Unitree G1

Prerequisites

Complete the following three steps before deploying for the first time. See the appendix at the end of this page for detailed instructions:

Tools
Flashing the SSD and Robot Firmware
System Configuration

Download ioenv

git clone https://github.com/ioai-tech/ioenv_cli.git

Load environment variables

echo 'source <path_to_ioenv_cli>/ioenv_cli/ioenv.sh' >> ~/.bashrc # Replace <path_to_ioenv_cli> with the local path to ioenv_cli
source ~/.bashrc

List available images with ioenv list

ioenv list

ENVIRONMENT          IMAGE      CONTAINER    STATUS    
-----------          -----      ---------    ------    
isaaclab             yes        --           --        
mujocosim            --         --           --        
onboard              --         --           --        
sonic_onboard        --         --           --        
sonic_sim            --         --           --   
sonic_train          yes        yes          running   

Pull an image with ioenv pull <ENVIRONMENT>

ioenv pull sonic_onboard

# Mirror download for users in regions with limited Docker Hub access
ioenv pull sonic_onboard --from-mirror

Running

# First terminal
ioenv run sonic_onboard

cd gear_sonic_deploy
source scripts/setup_env.sh
./deploy.sh real --input-type zmq_manager

# Second terminal
ioenv run sonic_onboard

source .venv_teleop/bin/activate
python gear_sonic/scripts/pico_manager_thread_server.py --manager

Training Sonic on a PC

note

The example commands train on an RTX 3090
Our actual training runs use 8x RTX PRO 6000 GPUs

Install Dependencies

Install Docker Tutorial

tip

You need to add your user to the docker group.

sudo usermod -aG docker $USER # Reboot required after running this

Install NVIDIA Container Toolkit Tutorial

Download ioenv

git clone https://github.com/ioai-tech/ioenv_cli.git

Load environment variables

echo 'source <path_to_ioenv_cli>/ioenv_cli/ioenv.sh' >> ~/.bashrc # Replace <path_to_ioenv_cli> with the local path to ioenv_cli
source ~/.bashrc

List available images with ioenv list

ioenv list

ENVIRONMENT          IMAGE      CONTAINER    STATUS    
-----------          -----      ---------    ------    
isaaclab             yes        --           --        
mujocosim            --         --           --        
onboard              --         --           --        
sonic_onboard        --         --           --        
sonic_sim            --         --           --   
sonic_train          yes        yes          running   

Pull an image with ioenv pull <ENVIRONMENT>

ioenv pull sonic_train

# Mirror download for users in regions with limited Docker Hub access
ioenv pull sonic_train --from-mirror

Data Preparation

System Requirements

GPU: NVIDIA GPU (L40 recommended)
CUDA: CUDA 12.x

Enter the sonic_train environment

ioenv run sonic_train

Optional: create symlinks under the mounted directory so data is not lost when the container restarts

mkdir -p /root/workspace/data
ln -s /root/workspace/data /opt/GR00T-WholeBodyControl/data

mkdir -p /root/workspace/sonic_release
ln -s /root/workspace/sonic_release /opt/GR00T-WholeBodyControl/sonic_release

Download model checkpoints and SMPL motion data from Hugging Face

cd /opt/GR00T-WholeBodyControl/

## A proxy is required here
export https_proxy=http://127.0.0.1:7897/
export http_proxy=http://127.0.0.1:7897/

python download_from_hf.py --training

Download the BONES-SEED dataset
Log in to Hugging Face, click your avatar in the top-right corner → Access Tokens, generate a token, and grant it the permission Read access to contents of all public gated repos you can access.
Fill out the form on the bones-seed Hugging Face page to gain download access: https://huggingface.co/datasets/bones-studio/seed

## Proxy required
export https_proxy=http://127.0.0.1:7897/
export http_proxy=http://127.0.0.1:7897/

# Login first
huggingface-cli login

## Download all files in bones-seed
huggingface-cli download bones-studio/seed --repo-type dataset --local-dir ./data/bones-seed

## We only need g1.tar.gz
export HF_HUB_ENABLE_HF_TRANSFER=1
huggingface-cli download bones-studio/seed g1.tar.gz --repo-type dataset --local-dir ./data/bones-seed

## Extract
tar -xzf data/bones-seed/g1.tar.gz -C data/bones-seed/g1

Convert the motion data

SONIC expects motion data in motion_lib PKL format. This step converts the Bones-SEED CSV files to PKL.

cd /opt/GR00T-WholeBodyControl/

python gear_sonic/data_process/convert_soma_csv_to_motion_lib.py \
    --input data/bones-seed/g1/csv/ \
    --output data/motion_lib_bones_seed/robot \
    --fps 30 \
    --fps_source 120 \
    --individual \
    --num_workers 16

Filter the data

Remove motions the G1 robot cannot perform (interactions with furniture, vehicles, stunt motions, raised platforms / elevated surfaces).

cd /opt/GR00T-WholeBodyControl/

python gear_sonic/data_process/filter_and_copy_bones_data.py \
    --source data/motion_lib_bones_seed/robot \
    --dest data/motion_lib_bones_seed/robot_filtered \
    --workers 16

File Structure

/opt/GR00T-WholeBodyControl/
├── data/motion_lib_bones_seed/
│   ├── robot/              # Full motion library (142K PKLs)
│   └── robot_filtered/     # Filtered subset (~130K PKLs)
└── gear_sonic/

Training

Before starting training, validate the full pipeline with sample_data

# Download sample data
python download_from_hf.py --sample

python gear_sonic/train_agent_trl.py \
    +exp=manager/universal_token/all_modes/sonic_release \
    num_envs=16 headless=True \
    ++manager_env.commands.motion.motion_lib_cfg.motion_file=sample_data/robot_filtered \
    ++manager_env.commands.motion.motion_lib_cfg.smpl_motion_file=sample_data/smpl_filtered

Then train with the converted BONES-SEED dataset

python gear_sonic/train_agent_trl.py \
    +exp=manager/universal_token/all_modes/sonic_release \
    num_envs=4096 headless=True \
    ++manager_env.commands.motion.motion_lib_cfg.motion_file=data/motion_lib_bones_seed/robot_filtered \
    ++manager_env.commands.motion.motion_lib_cfg.smpl_motion_file=data/smpl_filtered

Or fine-tune from a downloaded checkpoint

python gear_sonic/train_agent_trl.py \
    +exp=manager/universal_token/all_modes/sonic_release \
    +checkpoint=sonic_release/last.pt \
    num_envs=4096 headless=True \
    ++manager_env.commands.motion.motion_lib_cfg.motion_file=data/motion_lib_bones_seed/robot_filtered \
    ++manager_env.commands.motion.motion_lib_cfg.smpl_motion_file=data/smpl_filtered

Training checkpoints are saved under /opt/GR00T-WholeBodyControl/logs_rl/TRL_G1_Track/manager/universal_token/all_modes/sonic_release_test-<timestamp>/.

Evaluating a Checkpoint

You can evaluate using either metrics or rendered videos.

# --- Metrics ---
python gear_sonic/eval_agent_trl.py \
    +checkpoint=<path_to_checkpoint.pt> \
    +headless=True \
    ++eval_callbacks=im_eval \
    ++run_eval_loop=False \
    ++num_envs=128 \
    "+manager_env/terminations=tracking/eval" \
    "++manager_env.commands.motion.motion_lib_cfg.max_unique_motions=512"

# --- Render videos ---
python gear_sonic/eval_agent_trl.py \
    +checkpoint=<path_to_checkpoint.pt> \
    +headless=True \
    ++eval_callbacks=im_eval \
    ++run_eval_loop=False \
    ++num_envs=8 \
    ++manager_env.config.render_results=True \
    "++manager_env.config.save_rendering_dir=/tmp/renders" \
    ++manager_env.config.env_spacing=10.0 \
    "~manager_env/recorders=empty" "+manager_env/recorders=render"

If you cannot log in to wandb, use local mode to view training curves

WANDB_MODE=offline python gear_sonic/train_agent_trl.py ...   # offline mode
    wandb.wandb_project=my_project wandb.wandb_entity=my_team  # custom project
    use_wandb=false                                             # disable entirely

Exporting an ONNX Model

python gear_sonic/eval_agent_trl.py \
    +checkpoint=<path_to_checkpoint.pt> \
    +headless=True ++num_envs=1 \
    +export_onnx_only=true

The exported ONNX models are placed under /opt/GR00T-WholeBodyControl/logs_rl/TRL_G1_Track/manager/universal_token/all_modes/sonic_release_test-<timestamp>/exported/. You'll need the model_step_<num_step>_encoder.onnx and model_step_<num_step>_decoder.onnx files from that directory.

Deploying a Model You Trained Yourself

warning

As of 2026-05-21, deploying a self-trained ONNX file requires downloading g1_wrist_joints_10_clean.yaml, as noted by the author in this GitHub issue.

Deploying Your Model in Simulation

Make sure ioenv is installed and the sonic_onboard image has been pulled.
Copy the .onnx and .yaml files to the shared folder.

mkdir -p ~/ioenv_ws/sonic_sim_ws/sonic_patch
cp g1_wrist_joints_10_clean.yaml ~/ioenv_ws/sonic_sim_ws/sonic_patch
cp model_step_<num_step>_encoder.onnx ~/ioenv_ws/sonic_sim_ws/sonic_patch
cp model_step_<num_step>_decoder.onnx ~/ioenv_ws/sonic_sim_ws/sonic_patch

# First terminal
ioenv run sonic_sim

cd /opt/GR00T-WholeBodyControl
source .venv_teleop/bin/activate
python gear_sonic/scripts/run_sim_loop.py

# Second terminal
ioenv run sonic_sim

## Run with your own trained model
just run g1_deploy_onnx_ref lo model_step_<num_step>_decoder.onnx reference/example/ \  --obs-config /root/workspace/sonic_patch/g1_wrist_joints_10_clean.yaml \  --encoder-file /root/workspace/sonic_patch/model_step_<num_step>_encoder.onnx \  --planner-file planner/target_vel/V2/planner_sonic.onnx \  --input-type manager \  --enable-motion-recording \  --enable-csv-logs \  --disable-crc-check

# Third terminal
ioenv run sonic_sim

cd /opt/GR00T-WholeBodyControl
source .venv_teleop/bin/activate
python gear_sonic/scripts/pico_manager_thread_server.py --manager

Deploying Your Model on the Unitree G1

Make sure ioenv is installed and the sonic_onboard image has been pulled.
Copy the .onnx and .yaml files to the shared folder.

mkdir -p ~/ioenv_ws/sonic_onboard_ws/sonic_patch
cp g1_wrist_joints_10_clean.yaml ~/ioenv_ws/sonic_sim_ws/sonic_patch
cp model_step_<num_step>_encoder.onnx ~/ioenv_ws/sonic_sim_ws/sonic_patch
cp model_step_<num_step>_decoder.onnx ~/ioenv_ws/sonic_sim_ws/sonic_patch

# First terminal
ioenv run sonic_onboard

# Run with your own trained model
just run g1_deploy_onnx_ref enP8p1s0 /root/workspace/sonic_patch/model_step_<num_step>_decoder.onnx reference/example/ \  --obs-config /root/workspace/sonic_patch/g1_wrist_joints_10_clean.yaml \  --encoder-file /root/workspace/sonic_patch/model_step_<num_step>_encoder.onnx \  --planner-file planner/target_vel/V2/planner_sonic.onnx \  --input-type zmq_manager \  --enable-motion-recording \  --enable-csv-logs

# Second terminal
ioenv run sonic_onboard

source .venv_teleop/bin/activate
python gear_sonic/scripts/pico_manager_thread_server.py --manager

Appendix

Tools

An Ubuntu 22.04 laptop
An external drive enclosure, an SSD (256 GB or larger), and a data cable
Unitree G1
Tools (top to bottom): hex wrench, Torx (hex star) screwdriver, Phillips screwdriver

Flashing the SSD and Robot Firmware

Flash the Jetson Image to the SSD

Insert the SSD into the enclosure and connect it to your laptop.
Unmount the SSD.
```
sudo umount /dev/sda*
```
Format the SSD as ext4.
```
sudo mkfs -t ext4 /dev/sda
```
Download the Unitree G1 Jetson 6.2 image g1-nx-j6.2.img.bz2 from this link and flash it to the SSD. About 250 GB of data will be written.
```
bzip2 -dc g1-nx-j6.2.img.bz2 | sudo dd of=/dev/sda bs=4M status=progress conv=fsync
```

Eject the SSD when done.

sudo sync
sudo udisksctl power-off -b /dev/sda

Remove the SSD from the Unitree G1

Use the hex wrench to remove the screws around the rear handle.

Pry off the silver back cover.

Use the Torx screwdriver to remove the screws inside the red circles.

Peel off the inner black plastic back cover.

Use the Phillips screwdriver to unscrew the screws securing the SSD.

warning

The first time you remove the SSD screws, sealing adhesive is present. Align a smaller screwdriver with the screw and tap on top of the screwdriver with a hard object to break the seal.

Flash Jetpack 6.2 Firmware on the Unitree G1

Press the power button briefly, then press and hold it to power on the G1. Wait until all three power indicators are steadily lit.
Connect the robot to your laptop with a USB-C cable.
Press and hold the two white buttons to the right of the green LED simultaneously for 2 seconds.
While continuing to hold them, release the top white button (PWR) and keep holding the bottom button (REC) for another 2 seconds, until three green LEDs become two.

When only two LEDs are lit, the robot is in flashing mode. Open a new terminal on your computer and run lsusb. You should see an entry containing NVIDIA Corp. APX.
```
lsusb
```
Download Jetpack_6.2_nx.tar.bz2 from this link and run the following commands to flash the Jetson Orin NX Jetpack 6.2 firmware (about 8 minutes).
```
sudo tar -xjvf Jetpack_6.2_nx.tar.bz2
cd Jetpack_6.2_nx/Linux_for_Tegra
sudo ./flash_nx_module.sh
```
Once flashing is complete, reinstall the SSD and reattach the back covers.

System Configuration

Connect to the Unitree G1 over Wi-Fi

Connect the Unitree G1 to your laptop with an Ethernet cable.
SSH into the Unitree Orin NX development board. Wired interface address: 192.168.123.164, username unitree, password 123.
```
ssh unitree@192.168.123.164 # password 123
```
Enable Wi-Fi on the Unitree G1.
```
sudo nmtui
# Connect to Wi-Fi in the TUI
```
Check the IP address of the Unitree G1 wireless interface.
```
ifconfig
```
Once you have the wireless IP, you can disconnect the Ethernet cable and connect to the G1 over Wi-Fi.
```
ssh unitree@<WiFi-IP> # password 123
```

Enable High-Performance Mode on the Orin NX

On a freshly installed system, the Orin does not enable performance mode by default. When you check with jtop, CPU cores 5–8 are off, Jetson Clocks is inactive, and power is capped at 15W.

Enable high-performance mode

sudo nvpmodel -m 0

sudo jetson_clocks
sudo jetson_clocks --show

Run jtop again to confirm that high-performance mode is now active.

Install Jetpack Libraries

In jtop, switch to the INFO page — the CUDA libraries are missing and need to be installed manually.

Install jetpack via apt

sudo apt update
sudo apt install nvidia-jetpack

Once installed, you can check the library versions in jtop.

Expand the SSD Partition

After installation, df -h shows only 256 GB available, but the install was done on a 2 TB drive. Next, we'll expand the partition.

Check how much unused space the disk has.
```
sudo parted /dev/nvme0n1 print free
```
You can see 254 GB used (blue box) and 1744 GB unused (red box).

Append the unused space to the ext4 partition.

sudo parted /dev/nvme0n1
# Inside the parted CLI
resizepart 1 100%

# Online resize
sudo resize2fs /dev/nvme0n1p1

Verify — the ext4 partition has been expanded to 2 TB.
```
sudo parted /dev/nvme0n1 print free
```

Running the Sonic Simulation on a PC​

Install Dependencies​

Download ioenv​

Running​

Running XRoboToolkit on the Pico​

Installation​

Calibration​

Settings​

Deploying Sonic on Unitree G1​

Prerequisites​

Download ioenv​

Running​

Training Sonic on a PC​

Install Dependencies​

Download ioenv​

Data Preparation​

File Structure​

Training​

Evaluating a Checkpoint​

Exporting an ONNX Model​

Deploying a Model You Trained Yourself​

Deploying Your Model in Simulation​

Deploying Your Model on the Unitree G1​

Appendix​

Tools​

Flashing the SSD and Robot Firmware​

Flash the Jetson Image to the SSD​

Remove the SSD from the Unitree G1​

Flash Jetpack 6.2 Firmware on the Unitree G1​

System Configuration​

Connect to the Unitree G1 over Wi-Fi​

Enable High-Performance Mode on the Orin NX​

Install Jetpack Libraries​

Expand the SSD Partition​

Running the Sonic Simulation on a PC

Install Dependencies

Download ioenv

Running

Running XRoboToolkit on the Pico

Installation

Calibration

Settings

Deploying Sonic on Unitree G1

Prerequisites

Download ioenv

Running

Training Sonic on a PC

Install Dependencies

Download ioenv

Data Preparation

File Structure

Training

Evaluating a Checkpoint

Exporting an ONNX Model

Deploying a Model You Trained Yourself

Deploying Your Model in Simulation

Deploying Your Model on the Unitree G1

Appendix

Tools

Flashing the SSD and Robot Firmware

Flash the Jetson Image to the SSD

Remove the SSD from the Unitree G1

Flash Jetpack 6.2 Firmware on the Unitree G1

System Configuration

Connect to the Unitree G1 over Wi-Fi

Enable High-Performance Mode on the Orin NX

Install Jetpack Libraries

Expand the SSD Partition