Sim Data Export: HDF5 to LeRobot#

In this lesson, we’ll convert the teleoperated HDF5 demonstrations from the static apple-to-plate task into LeRobot format for GR00T 1.7 post-training.

To do that, we’ll:

Set the dataset directory inside the Arena container.
Verify the HDF5 conversion configuration.
Convert the recorded HDF5 dataset to LeRobot format.
Identify the resulting LeRobot dataset artifacts.

Convert to LeRobot Format#

GR00T 1.7 consumes datasets in LeRobot format. The conversion runs inside the standard Base Arena container.

Continue in the Arena container terminal from the last module.
Command for reference, if you need to re-attach or launch container.
./docker/run_docker.sh
Confirm DATASET_DIR is still set to /datasets/isaaclab_arena/static_apple_tutorial
```
echo $DATASET_DIR
```
If it’s not set, reset it:
```
export DATASET_DIR=/datasets/isaaclab_arena/static_apple_tutorial
```

Tip

At this point, you can either choose to use a dataset you recorded yourself, or use ours.

Use these commands below to download the released simulation dataset if you want to skip teleoperation.

Download our pre-recorded dataset, nvidia/Arena-G1-Static-PickNPlace-Task, from Hugging Face. Run these commands where DATASET_DIR points to the target directory.

If inside Docker, use /datasets/.... Hugging Face CLI is already installed.

If outside Docker, use the matching host path. The Hugging Face CLI from huggingface_hub must be installed in that environment.

hf download \
   nvidia/Arena-G1-Static-PickNPlace-Task \
   arena_g1_static_apple_dataset_recorded_200_demos.hdf5 \
   --repo-type dataset \
   --local-dir $DATASET_DIR

mv "$DATASET_DIR/arena_g1_static_apple_dataset_recorded_200_demos.hdf5" \
   "$DATASET_DIR/arena_g1_static_apple_dataset_recorded.hdf5"

The rename lets the default g1_static_apple_config.yaml use the released HDF5 without changing hdf5_name.

Caution

isaaclab_arena_gr00t/lerobot/convert_hdf5_to_lerobot.py expects each episode to include ego RGB under observations/camera_obs/robot_head_cam_rgb; see pov_cam_name_sim in the conversion config. Before bulk collection, run the conversion once on a short recording to confirm the layout matches.

Edit isaaclab_arena_gr00t/lerobot/config/g1_static_apple_config.yaml so that:

hdf5_name matches your recorded file - arena_g1_static_apple_dataset_recorded.hdf5 in this example.
data_root matches $DATASET_DIR

Still inside the container, convert the HDF5 dataset to LeRobot format.

python isaaclab_arena_gr00t/lerobot/convert_hdf5_to_lerobot.py \
  --yaml_file isaaclab_arena_gr00t/lerobot/config/g1_static_apple_config.yaml

This creates a folder at $DATASET_DIR/arena_g1_static_apple_dataset_recorded/lerobot containing parquet files with states and actions, MP4 camera recordings, and dataset metadata.

Tip

One option for sharing or post-training on cloud infrastructure, is to upload the LeRobot folder to Hugging Face Hub. Authenticate with hf auth login first, then replace <your-username> with your Hugging Face username or organization. You can upload all checkpoints, or just the latest one.

hf upload \
  <your-username>/my_apple_sim_datasets \
  $DATASET_DIR/arena_g1_static_apple_dataset_recorded/lerobot \
  . \
  --repo-type dataset

Review Conversion Configuration#

The converter is controlled by isaaclab_arena_gr00t/lerobot/config/g1_static_apple_config.yaml.

Below is an example file:

The main differences from the loco-manipulation box config, g1_locomanip_config.yaml, are the data_root and hdf5_name values pointing at the static apple-to-plate dataset and the language_instruction for the static apple-to-plate task. The 43-DoF action layout, embodiment tag, modality template, and joint-space configurations are shared with the loco-manipulation variant. The static workflow does not need its own GR00T embodiment config because the upper-body action channels and observation modalities are identical; only the recorded body channel stays at zero throughout each demo.

Note

The recorder’s processed_actions field already contains the 43-DoF joint-space targets that PinkIK produced during teleoperation. That is why this workflow records with g1_wbc_agile_pink and evaluates with g1_wbc_agile_joint: the policy never sees the end-effector pose targets PinkIK consumed. It only sees the joint targets PinkIK produced.

Key Takeaways#

We converted the recorded static apple-to-plate HDF5 demonstrations into LeRobot format. The resulting dataset is ready for GR00T 1.7 post-training in the standalone Isaac-GR00T checkout.