sadpiggy/rtr_robot_sys_zarr

RTR Robot Sys — Zarr Training Datasets

Training datasets for the paper "Learning High-Frequency Continuous Action Chunks in Latent Space", used to train the latent-space high-frequency policies and the Reuse-then-Refine (RTR) chunk-refinement strategy on three real-world contact-rich tasks.

• Paper: Learning High-Frequency Continuous Action Chunks in Latent Space
• Project page: sjtu-zhao-lab.github.io/RTR
• Code: github.com/tars-robotics/RTR

The training-side documentation, including the full unpacking + training pipeline, lives in the code repo at docs/best_practice/train/download_data.md. End users should follow that guide — this card is just a short overview of the released files.

If you want to train a LeRobot-style policy (e.g. pi0.5) on this data, we also ship a script that converts these zarr archives into the LeRobot dataset format. See docs/best_practice/train/lerobot.md in the code repo for instructions.

Contents

Each task is shipped as a single .tar archive that contains two top-level entries:

<task>.tar
├── rdp_zarr/
│   └── replay_buffer.zarr/      # zarr group with episode data
└── rdp_pca/
    ├── pca_matrix1.npy
    ├── pca_matrix2.npy
    ├── pca_mean1.npy
    └── pca_mean2.npy

We use uncompressed .tar because the zarr replay buffers are already compressed at the chunk level — additional gzip / zstd adds cost with negligible savings. One large file per task is also the most LFS-friendly upload pattern on the Hub and downloads cleanly resume.

File	Frequency	rdp_zarr size	rdp_pca size
peel_cucumber_15hz.tar	15 Hz	3.7 GB	< 1 MB
peel_cucumber_60hz.tar	60 Hz	15 GB	< 1 MB
wipe_vase_15hz.tar	15 Hz	2.6 GB	< 1 MB
wipe_vase_60hz.tar	60 Hz	9.7 GB	< 1 MB
write_board_15hz.tar	15 Hz	3.2 GB	< 1 MB
write_board_60hz.tar	60 Hz	12 GB	< 1 MB
Total	—	~46 GB	~2 MB

Quick start

1. Download

Download every .tar from the repo into a local directory:

pip install -U "huggingface_hub[cli]"

huggingface-cli download sadpiggy/rtr_robot_sys_zarr \
    --repo-type dataset \
    --local-dir data/zarr_dataset \
    --local-dir-use-symlinks False

To grab just one task:

huggingface-cli download sadpiggy/rtr_robot_sys_zarr \
    --repo-type dataset \
    --include "peel_cucumber_60hz.tar" \
    --local-dir data/zarr_dataset \
    --local-dir-use-symlinks False

Python alternative:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="sadpiggy/rtr_robot_sys_zarr",
    repo_type="dataset",
    local_dir="data/zarr_dataset",
    local_dir_use_symlinks=False,
    allow_patterns=["*.tar"],
)

2. Unpack into the layout expected by training

The training scripts in the code repo read from data/ckpts/<task>/, so each .tar should be extracted into the matching task directory:

mkdir -p data/ckpts
for tar in data/zarr_dataset/*.tar; do
    task=$(basename "$tar" .tar)
    mkdir -p "data/ckpts/$task"
    tar -xf "$tar" -C "data/ckpts/$task"
done

After extraction data/ckpts/<task>/ will contain rdp_zarr/ and rdp_pca/ ready to train.

3. Sanity check

import zarr
from pathlib import Path

for task in sorted(Path("data/ckpts").iterdir()):
    zarr_root = task / "rdp_zarr" / "replay_buffer.zarr"
    if not zarr_root.exists():
        continue
    root = zarr.open(str(zarr_root), mode="r")
    n_episodes = root["meta/episode_ends"].shape[0]
    n_steps = int(root["meta/episode_ends"][-1]) if n_episodes else 0
    print(f"{task.name:24s}  episodes={n_episodes:4d}  steps={n_steps}")

License

Released under the MIT License, matching the code repository.

Citation

If you use this dataset, please cite:

@article{wang2026learning,
  title={Learning High-Frequency Continuous Action Chunks in Latent Space},
  author={Wang, Kunyun and Zheng, Yuhang and Zheng, Yupeng and Zhao, Jieru and Ding, Wenchao},
  journal={arXiv preprint arXiv:2605.24931},
  year={2026}
}