TexasPokerRobot

TexasPokerRobot is a robot manipulation dataset collected in a Texas poker tabletop environment. The raw episodes are stored as compressed NumPy .npz files, organized by action folder. This release adds a Hugging Face-compatible manifest at data/train.csv so the dataset has a standard loadable split and a working Dataset Viewer while preserving the original raw episode files.

Dataset Summary

• 1,470 raw episode files
• 14 action folders, with 105 episodes per action
• 377.94 GB of raw .npz episode data
• RGB observations from three cameras, depth observations from three cameras, and robot joint state streams
• License: CC BY 4.0

Actions

Action	Episodes	Raw size
pick_up_left	105	40.19 GB
pick_up_right	105	29.63 GB
pull_5	105	29.69 GB
pull_10	105	29.00 GB
pull_50	105	31.83 GB
pull_100	105	31.17 GB
push_5	105	28.44 GB
push_10	105	33.60 GB
push_50	105	35.04 GB
push_100	105	31.10 GB
put_down_left	105	18.23 GB
put_down_right	105	14.59 GB
show_left	105	13.23 GB
show_right	105	12.19 GB

Manifest Fields

The train split is a manifest. Each row points to one raw episode file.

Column	Description
action	Action folder name.
episode_index	Episode number within the action folder.
file_path	Path to the raw .npz file in this dataset repository.
file_name	Raw episode filename.
format	Raw file format, currently npz.
size_bytes	Size of the raw episode file.
sha256	LFS SHA-256 checksum for the raw episode file.
download_url	Direct Hub URL for the raw episode file.
modalities	Modalities present in the raw episode.
num_rgb_cameras	Number of RGB camera streams.
num_depth_cameras	Number of depth camera streams.

Loading

Load the manifest with the datasets library:

from datasets import load_dataset

ds = load_dataset("Winniechen2002/TexasPokerRobot")
row = ds["train"][0]
print(row["action"], row["file_path"])

Download and inspect a raw episode:

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Winniechen2002/TexasPokerRobot",
    filename=row["file_path"],
    repo_type="dataset",
)

episode = np.load(path, allow_pickle=True)
print(episode.files)

Some robot state arrays are stored as NumPy object arrays, so loading a raw episode requires allow_pickle=True. Only use this option with dataset files you trust.

Intended Uses

This dataset is intended for robot learning research, imitation learning, action-conditioned perception, manipulation policy development, and analysis of tabletop robot trajectories.

Limitations

The Dataset Viewer previews the manifest rows, not the full RGB, depth, and joint-state tensors. The raw episode files are large, and no official train/validation/test benchmark split is provided beyond the manifest split.