This benchmark tests loco-manipulation capabilities on the Unitree G1 humanoid robot using whole-body control in MuJoCo. Tasks require navigation, object manipulation, and maintaining balance across the entire body.
Benchmark results
Checkpoint: nvidia/GR00T-N1.6-G1-PnPAppleToPlate
| Task | Success rate |
|---|
| PnPAppleToPlate | 58% |
This task exhibits high evaluation variance. Fluctuations of ±15% are expected across different evaluation runs.
Fine-tuning
You can skip this section and directly evaluate with the provided checkpoint above.
Clone dataset repository
Clone without downloading all files:cd examples/GR00T-WholeBodyControl
git clone --filter=blob:none --no-checkout \
https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim
cd PhysicalAI-Robotics-GR00T-X-Embodiment-Sim
Setup sparse checkout
Download only the G1 dataset:# Initialize sparse-checkout
git sparse-checkout init --cone
# Select folder to download
git sparse-checkout set unitree_g1.LMPnPAppleToPlateDC
# Checkout files
git checkout
# Pull LFS files
git lfs pull
Run fine-tuning
cd ../../../
uv run bash examples/GR00T-WholeBodyControl/finetune_g1.sh
Evaluation
Setup environment
Install the required dependencies (only needs to be done once):
apt-get update
apt-get install libegl1-mesa-dev libglu1-mesa
bash gr00t/eval/sim/GR00T-WholeBodyControl/setup_GR00T_WholeBodyControl.sh
Run evaluation
Start policy server
In Terminal 1, choose one of the following options:Option 1: Local fine-tuned checkpointuv run python gr00t/eval/run_gr00t_server.py \
--model-path /tmp/g1_finetune/checkpoint-10000/ \
--embodiment-tag UNITREE_G1 \
--use-sim-policy-wrapper
uv run python gr00t/eval/run_gr00t_server.py \
--model-path nvidia/GR00T-N1.6-G1-PnPAppleToPlate \
--embodiment-tag UNITREE_G1 \
--use-sim-policy-wrapper
Start evaluation client
In Terminal 2:gr00t/eval/sim/GR00T-WholeBodyControl/GR00T-WholeBodyControl_uv/.venv/bin/python \
gr00t/eval/rollout_policy.py \
--n_episodes 10 \
--max_episode_steps=1440 \
--env_name gr00tlocomanip_g1_sim/LMPnPAppleToPlateDC_G1_gear_wbc \
--n_action_steps 20 \
--n_envs 5
Available tasks
gr00tlocomanip_g1_sim/LMPnPAppleToPlateDC_G1_gear_wbc - Pick apple and place on plate with dynamic constraints
Real robot deployment
When working with real Unitree G1 robot data, you have two options:
Option 1: Using GR00T-WholeBodyControl
If you collected data using GR00T-WholeBodyControl, leverage the UNITREE_G1 embodiment tag. This is a pre-trained embodiment with models already trained on in-the-wild Unitree G1 datasets.
Option 2: Custom whole-body controller
If your data was collected using a different whole-body controller, create and fine-tune with a NEW_EMBODIMENT tag. This allows you to define a custom embodiment tailored to your specific controller setup.
See the fine-tune new embodiment guide for detailed instructions.