smilegeng/stanford_hydra_dataset

Stanford HYDRA (LeRobot) — TsFile

This dataset converts the numeric time-series of the HuggingFace LeRobot dataset lerobot/stanford_hydra_dataset losslessly to the Apache TsFile format, and keeps the original camera videos alongside.

Original dataset

• Source dataset: lerobot/stanford_hydra_dataset
• Format: LeRobot v3
• License: MIT
• Content: a robot-manipulation dataset — 570 episodes / 358,234 frames / 10 fps, with proprioceptive state, actions, rewards, and two RGB camera streams.

What is in this repository

data/
└── stanford_hydra.tsfile          # numeric time-series (converted)
videos/
├── observation.images.image/...   # main camera (MP4, copied verbatim)
└── observation.images.wrist_image/...  # wrist camera (MP4, copied verbatim)

• data/stanford_hydra.tsfile — the converted numeric time-series (state / action / reward / indices).
• videos/… — the two camera MP4 streams, copied verbatim from the source (not converted; TsFile is a numeric time-series format and does not store video).

TsFile storage mapping (table model)

Role	Column(s)	Type	Notes
TAG	episode_id	STRING	episode_{episode_index}, 570 devices (one per episode)
Time	frame_index * 100 ms	INT64 (ms)	10 fps; frame_index restarts at 0 each episode and is strictly increasing
FIELD	state_0 … state_7	FLOAT	observation.state[8] expanded
FIELD	action_0 … action_6	FLOAT	action[7] expanded
FIELD	episode_index, frame_index, sample_index, task_index	INT64	indices (index → sample_index)
FIELD	episode_timestamp_s, next_reward	FLOAT	(timestamp → episode_timestamp_s, next.reward → next_reward)
FIELD	next_done	BOOLEAN	(next.done)

Conversion notes

• Only the numeric time-series is converted. The observation.images.* features are dtype=video in LeRobot — their pixels live in the MP4 files under videos/, which are kept verbatim here.
• TAG = episode_id (570 devices). Time = frame_index × 100 ms. Because frame_index restarts at 0 within each episode and is strictly increasing, every device's time axis is strictly increasing — no de-duplication or offset needed.
• Array columns expanded: observation.state[8] → state_0..state_7, action[7] → action_0..action_6 (all FLOAT, matching the source float32).
• Column names with dots are made TsFile-safe (next.reward → next_reward, …).
• No columns dropped, no rows dropped: all 358,234 frames preserved.

Usage

from tsfile import TsFileReader

reader = TsFileReader("data/stanford_hydra.tsfile")
schemas = reader.get_all_table_schemas()
tname = next(iter(schemas))

cols = ["episode_id", "state_0", "action_0", "next_reward", "task_index"]
with reader.query_table(tname, cols, batch_size=65536) as rs:
    while (batch := rs.read_arrow_batch()) is not None:
        df = batch.to_pandas()
        # ... process ...
reader.close()

Citation

@misc{stanford_hydra_lerobot,
  title  = {Stanford HYDRA dataset (LeRobot)},
  author = {LeRobot team},
  url    = {https://huggingface.co/datasets/lerobot/stanford_hydra_dataset},
  publisher = {Hugging Face}
}

Original dataset licensed under MIT.