EgoVerse: Egocentric Data for Robot Learning from Around the World

This repository contains the data processing, training and evaluation code for EgoVerse.

• Camera intrinsics are now MANDATORY in every episode's zarr.json, stored as a {camera_key: 3×4 K} dict (single-camera = one entry, e.g. {"front_1": K}). ZarrWriter.create_and_write raises if it is missing or not a non-empty dict. extrinsics (robots) is now strictly None or a non-empty dict.
• Embodiments collapsed: all human demonstration data is a single human_* embodiment (human_right_arm/human_left_arm/human_bimanual, ids 1–3); the robot Eva is eva_* (ids 4–6). Vendor labels (aria_*, mecka_*, scale_*, lightwheel_*) are removed at the embodiment level — the data source now lives only in the SQL lab field. Conversion scripts write human_*.
• Aria processing — EE/wrist orientation fixed: aria EE-pose and wrist-pose orientations are now correct and consistent across both hands — the left and right hand share the same canonical T_ROT_CAM convention, aligned with the robot (Eva) tool frame, so human and robot EE orientations line up for joint training / visualization. Re-process aria data to pick up the corrected orientations.
• ⚠️ Action required — RE-DOWNLOAD your data. The embodiment string is stored inside each episode's zarr.json, and the reader looks it up with no alias fallback (get_embodiment_id): a locally cached episode written before the collapse still says aria_bimanual / scale_bimanual / mecka_bimanual, and loading it now hard-crashes with KeyError: 'ARIA_BIMANUAL'. Delete any local cache and re-pull the reprocessed episodes (they carry human_* + the mandatory intrinsics). Data producers: re-process and re-upload so zarr.json includes intrinsics — see CONTRIBUTING_DATA.md. Mecka and Scale will be asked to add intrinsics to their exports.

Mecka removed some poorer quality episodes and replaced them with higher quality alternatives.

The Scale dataset was fully reprocessed on 05/03/2026. All active Scale episodes now use newly generated episode hashes, Zarr paths, and preview MP4 paths. If you previously referenced Scale episode hashes from an older export or intermediate processing run, refresh from the SQL episode table before downloading or training. Old Scale hashes should be treated as stale and should not be mixed with the current active dataset.

• egomimic/trainHydra.py: Main training script, powered by Pytorch Lightning and Hydra (DDP supported)
• egomimic/hydra_configs: Train configs for each algorithm
• egomimic/algo: Algorithm code: EgoMimic (HPT based), Pi
• egomimic/scripts/aloha_process: Process raw aloha hdf5 to zarr/lerobot
• egomimic/scripts/aria_process: Process aria vrs to zarr/lerobot

if uv not installed

curl -LsSf https://astral.sh/uv/install.sh | env UV_INSTALL_DIR="/path/to/flash/storage" sh

git clone git@github.com:GaTech-RL2/EgoVerse.git
cd EgoVerse
uv venv emimic --python 3.11
source emimic/bin/activate
uv pip install -e .
uv run pre-commit install

pyproject.toml (locked in uv.lock) is the dependency spec for the package. Three things live outside it: requirements.txt (used only by the Eva robot Dockerfile), requirements-ray.txt (the Ray data-processing image) and scale_sensor_fusion_io, needed only by the Scale converter under external/scale/scripts (uv pip install scale-sensor-fusion-io).

For the Pi0.5 policy also install the openpi fork from the submodule. It is installed without its own dependencies (its lerobot pin needs av>=14.2, which conflicts with our av==12.0.0), then its transformers patch is copied in:

git submodule update --init external/openpi
uv pip install --no-deps -e external/openpi
cp -r external/openpi/src/openpi/models_pytorch/transformers_replace/* emimic/lib/python3.11/site-packages/transformers/

Re-run the cp after any uv sync or transformers reinstall, which silently undoes the patch. See pi05.md for details.

git clone --recursive git@github.com:GaTech-RL2/EgoVerse.git
cd EgoVerse
conda create -n emimic python=3.11
conda activate emimic
pip install -e .
pre-commit install

Download the AWS cli

 curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "awscliv2.zip"
 unzip awscliv2.zip
 ./aws/install -i ~/aws-cli -b ~/bin

Set up your AWS keys to access our cloud storage

aws configure
AccessKeyId: AKIAYDKH4BNCAYHE5NG2
SecretAccessKey: rGjT6NSh55YiB9MC9EyNGpVy8qcaTn4i19OmkhRW
Default region name: us-east-2
Default output format:
./egomimic/utils/aws/setup_secret.sh

setup_secret.sh will allow your current env to download data from cloudflare.

Every cluster-specific default in the configs can be overridden without editing yaml:

Set git config --global submodule.recurse true if you want git pull to automatically update the submodule as well. Set your wandb project in egomimic/hydra_configs/logger/wandb.yaml

For the integrated hydra submitit plugin to work, make the following modification...

/path/to/your/venv/emimic/lib/python3.11/site-packages/hydra_plugins/hydra_submitit_launcher/submitit_launcher.py

Change line 144 to

        jobs = executor.map_array(self, *zip(*job_params))

        return [asyncLauncher() for j in jobs]

class asyncLauncher:
    def __init__(self):
        self.return_value = 0

I wanted to package this change nicely, but the hydra package is built very weirdly.

Visit https://partners.mecka.ai/egoverse to view our entire dataset in the web!

To visualize data programatically see zarr_data_viz.ipynb

To programatically view the SQL table of all episodes + metadata see sql_tutorial.ipynb

latent_inspector.py also ships a local web app for browsing a folder of per-episode zarrs — scrub any episode, overlay the recorded actions (cartesian trajectory / orientation axes / MANO keypoints), and toggle language annotations. Frames are rendered server-side using each episode's zarr.json camera intrinsics, so it works for any embodiment (the overlay is drawn by that episode's embodiment class, e.g. Human/Eva; human poses are projected in the head frame).

python egomimic/scripts/data_visualization/latent_inspector.py \
    --dataset-path /path/to/folder_of_zarrs \
    --host 127.0.0.1 --port 8050
# then open http://localhost:8050

--dataset-path is a directory of <episode>.zarr stores (each with images.front_1, left/right.obs_ee_pose, obs_head_pose, optional *.obs_keypoints and annotations, and intrinsics in its zarr.json). In the browser: pick an episode (searchable by filename or annotation text), scrub the frame slider or press ▶ to play, choose an overlay (None / Cartesian / Orientation / Keypoints), and toggle annotations on/off.

To browse data on a remote machine, run the app there and forward the port — ssh -L 8050:<node>:8050 <host> — then open http://localhost:8050 (rendering locally is far more responsive than over the tunnel).

While our training pipeline automatically downloads data, you can manually download data via sync_s3.py

For example, to download all our flagship Aria fold clothes data...

python egomimic/scripts/data_download/sync_s3.py \
     --local-dir <local directory> \
     --filters aria-fold-clothes

Basic training run (robot BC)...

python egomimic/trainHydra.py --config-name=train_zarr_cartesian

For full instructions on training see training.md

See embodiment_tutorial.ipynb as reference to write a conversion script for your own data.