Unitree G1 Policy Training#
Fine-tuning adapts GR00T N1.7 to a LeRobot dataset of physical G1 demonstrations and saves checkpoints for LEAPP export.
Prerequisites#
Run on an x86_64 workstation with an L40-class GPU or better (40+ GB VRAM).
Copy the complete dataset from Data Export to this workstation. Use its absolute local path as
<dataset-dir>below.Choose an absolute
<output-dir>for the training run.Install
uvusing the host tools instructions.
Prepare GR00T N1.7#
Use the nvidia-isaac/gr00t-leapp-export repository and the gr00t_workflow_0.1 tag for fine-tuning a policy for Unitree G1 and downstream export compatibility.
Tip
If you want to skip fine-tuning, use the provided nvidia/GR00T-N1.7-ApplePnP-V1 real-robot model.
Continue to LEAPP Export with a local copy of the model
and its corresponding LeRobot dataset.
Clone
gr00t-leapp-export:git clone --branch gr00t_workflow_0.1 --recurse-submodules https://github.com/nvidia-isaac/gr00t-leapp-export.git
Enter the repository and install its environment in the same terminal:
cd gr00t-leapp-export
uv sync --python 3.10
Accept model access terms for
nvidia/GR00T-N1.7-3Bandnvidia/Cosmos-Reason2-2B, then authenticate:uv run hf auth login
Training Configuration#
Parameter |
Description |
Value |
|---|---|---|
|
Pretrained GR00T weights |
|
|
Path to real-world LeRobot dataset |
|
|
Robot embodiment identifier |
|
|
Total training steps |
|
|
Checkpoint frequency |
|
|
Effective batch size |
|
Validate the Dataset#
Before training, validate
<dataset-dir>/meta/info.json:fpsmatches the recordersync_rate(30 Hz by default).total_episodesandtotal_framesreflect the expected recording volume.
Use at least 200 successful episodes for the baseline apple-to-plate recipe, following the recording quality guidance.
Modality Config#
For datasets of Unitree G1 demonstrations, the converter writes a default modality config file:
<dataset-dir>/new_embodiment_config_defaults.py
Use this file directly in the fine-tuning command.
It includes:
Seven Unitree G1
statekeys fromleft_legthroughright_hand.RELATIVEaction representation for arms.ABSOLUTEaction representation for hands, waist, and optional locomotion or effort keys.The model-default 16-step action prediction horizon.
Per-body-part
effort_<group>action keys only when the recorder observed non-zero feed-forward torques.
Note
For the default apple-to-plate dataset, no additional modality edits are required.
Launch Training#
Run fine-tuning from the root of
gr00t-leapp-export:CUDA_HOME=/usr/local/cuda \ CUDA_VISIBLE_DEVICES=0 \ PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \ uv run python gr00t/experiment/launch_finetune.py \ --base-model-path nvidia/GR00T-N1.7-3B \ --dataset-path <dataset-dir> \ --embodiment-tag NEW_EMBODIMENT \ --modality-config-path <dataset-dir>/new_embodiment_config_defaults.py \ --num-gpus 1 \ --output-dir <output-dir> \ --max-steps 10000 \ --save-steps 2000 \ --global-batch-size 32 \ --dataloader-num-workers 4
The training script saves checkpoints to <output-dir>/checkpoint-<step>/.
The --max-steps, --save-steps, and --global-batch-size values are starting points. Tune them for dataset size and available hardware.
See also
Refer to the following resources for additional detail:
Next Stage#
Confirm that <output-dir>/checkpoint-<step>/ exists after training completes.
Use the chosen checkpoint directory as <checkpoint-dir> in
Unitree G1 LEAPP Export, together with the same
<dataset-dir> used for training.