G1 Policy Training#
Train a GR00T N1.7 policy on a converted Unitree G1 apple-to-plate dataset in NVIDIA Isaac GR00T’s separate environment. The output is a checkpoint for G1 policy evaluation.
Use your own dataset converted in the previous module, or download the training-ready dataset below. Both paths continue with training setup on this page; the provided dataset does not require the earlier data-preparation modules.
Use a Published Dataset#
We provide a training-ready LeRobot dataset of 2,500 generated G1 demonstrations. It is already converted: skip recording, data generation, and conversion and continue with training setup after downloading it.
The same Hugging Face repository also contains source HDF5 files. Those are
separate from the LeRobot dataset used here. The command below downloads only
the converted e2e_ra_g1_apple_to_plate_2500_generated_demos_lerobot directory.
Run it from the workflow repository root.
If you converted your own dataset in the previous module, skip this download and continue to G1 policy training setup below.
Before downloading, complete Hugging Face access setup.
uv run --frozen --no-sync hf download nvidia/GR00T-E2E-Reference-Workflow-SIM-G1-N1.7-AppletoPlate-Dataset \
--revision 36234f4b48dad27f218d8e29ec7710ad5a90b48c \
--repo-type dataset \
--include 'e2e_ra_g1_apple_to_plate_2500_generated_demos_lerobot/**' \
--local-dir outputs/g1/downloads
The dataset is saved in
outputs/g1/downloads/e2e_ra_g1_apple_to_plate_2500_generated_demos_lerobot,
which contains data/, videos/, and meta/. Select Provided dataset in
Training Inputs.
Continue with G1 policy training setup,
starting with GR00T preparation and model access. The demonstrations use G1 in Isaac Lab-Arena and include navigation and
base-height commands; demonstrations recorded with a fixed-base G1 omit these commands.
Prerequisites#
A downloaded LeRobot dataset above, or a validated dataset from G1 Dataset Conversion.
Data recorded with Pink IK and AGILE in either Isaac Lab-Arena or native Isaac Lab, including their AutoData outputs. Both supply the navigation and base-height commands required by the G1 modality. Native fixed-base data omits these fields and needs custom training and evaluation integration outside this recipe.
A training GPU with at least 48 GB VRAM. The documented single-GPU configuration uses an NVIDIA RTX 6000 Ada; 128 GB system RAM is recommended. The recorded estimate is about 2–3 hours for 20,000 steps on that GPU.
Authorized Hugging Face access to both
nvidia/GR00T-N1.7-3Bandnvidia/Cosmos-Reason2-2B. The training processor loads Cosmos as a separate dependency.
Run the preparation commands from this repository’s root. Training uses
a standalone Isaac-GR00T checkout, pinned by runtime.toml. For Unitree H2,
use H2 policy training.
Prepare NVIDIA Isaac GR00T in Your Workspace#
From this repository’s root, create the standalone checkout and its locked environment. Run these blocks in order in the same terminal:
export E2E_ROOT="$PWD" export ISAAC_GR00T_ROOT=/absolute/path/outside/this/repo/Isaac-GR00T
git clone https://github.com/NVIDIA/Isaac-GR00T.git "$ISAAC_GR00T_ROOT" git -C "$ISAAC_GR00T_ROOT" checkout 4b1dca9d88d2a0b9ea5a65aa61c82ff89f5c4f0e (cd "$ISAAC_GR00T_ROOT" && uv sync --frozen) "$E2E_ROOT/tools/gr00t_runtime.sh" \ --gr00t-root "$ISAAC_GR00T_ROOT" --check
Keep this checkout clean. For an existing checkout at that revision, sync its environment and run the same runtime check.
Model Access#
Login to HF and confirm authentication
"$ISAAC_GR00T_ROOT/.venv/bin/hf" auth login
"$ISAAC_GR00T_ROOT/.venv/bin/hf" auth whoami
Download the base model and verify Cosmos access before training:
"$ISAAC_GR00T_ROOT/.venv/bin/hf" download nvidia/GR00T-N1.7-3B
"$ISAAC_GR00T_ROOT/.venv/bin/hf" download nvidia/Cosmos-Reason2-2B config.json
Both commands must succeed with the credentials used for training. The processor downloads its remaining Cosmos files when training starts. A locally cached GR00T base model does not establish Cosmos access.
Training Inputs#
Open a fresh terminal outside any distributed training job before setting the paths. Set them in that terminal using absolute values:
export E2E_ROOT=/absolute/path/to/e2e-reference-app export ISAAC_GR00T_ROOT=/absolute/path/to/Isaac-GR00T export TRAIN_OUTPUT="$E2E_ROOT/outputs/g1/training" export MODALITY_CONFIG="$E2E_ROOT/src/e2e_robot_workflows/gr00t/g1_static_apple_modality_config.py"
Use a new training output directory for an independent run.
Choose one dataset using the tabs below. Run only the command for your chosen input:
Use the training-ready dataset downloaded above:
export TRAIN_DATASET="$E2E_ROOT/outputs/g1/downloads/e2e_ra_g1_apple_to_plate_2500_generated_demos_lerobot"
Use the dataset converted from your recordings:
export TRAIN_DATASET="$E2E_ROOT/outputs/g1/lerobot"
Use the converted AutoData output:
export TRAIN_DATASET="$E2E_ROOT/outputs/g1/lerobot-generated"
The fine-tuning command below uses your selected dataset.
Run Fine-Tuning#
In the same terminal where the training variables are set, launch one GPU directly through Python:
(cd "$ISAAC_GR00T_ROOT" && \ "$E2E_ROOT/tools/gr00t_runtime.sh" --gr00t-root "$ISAAC_GR00T_ROOT" -- \ python gr00t/experiment/launch_finetune.py \ --base-model-path nvidia/GR00T-N1.7-3B \ --dataset-path "$TRAIN_DATASET" \ --output-dir "$TRAIN_OUTPUT" \ --embodiment-tag new_embodiment \ --modality-config-path "$MODALITY_CONFIG" \ --global-batch-size 12 \ --max-steps 20000 \ --num-gpus 1 \ --save-steps 5000 \ --save-total-limit 5 \ --no-tune-llm \ --tune-visual \ --tune-projector \ --tune-diffusion-model \ --dataloader-num-workers 8 \ --color-jitter-params \ brightness 0.3 \ contrast 0.4 \ saturation 0.5 \ hue 0.08)
Important
Use the G1 modality with
--embodiment-tag new_embodiment. It specifies the 40-step horizon required by GR00T N1.7. The built-inunitree_g1_full_body_with_waist_height_nav_cmdembodiment option will fail.This configuration tunes the visual backbone, projector, and diffusion model while freezing the LLM. Reduce the batch size if GPU memory is insufficient.
Confirm that training completes with finite loss and writes
"$TRAIN_OUTPUT/checkpoint-20000". Intermediate checkpoints are saved every 5,000 steps. Keep the full checkpoint directory, including its processor, statistics, trainer state, and experiment configuration.
Next Stage#
Continue to G1 policy evaluation with
outputs/g1/training/checkpoint-20000. Its preflight validates the checkpoint’s
required files and modality before starting the server.