THULab/FT-LLM-2026-RAMEN_libero

FT-LLM-2026-RAMEN LIBERO TsFile

This repository is the Apache TsFile edition of FT-LLM-2026-RAMEN/libero, a LeRobot v3.0 Panda manipulation dataset based on the LIBERO benchmark. It contains numeric robot state, actions, frame timing, episode/task dimensions, and the source meta/ records.

Source and attribution

• Source dataset: FT-LLM-2026-RAMEN/libero
• Source organization: FT-LLM-2026-RAMEN
• Repository contributor: yusei926; the public history credits this account with the LeRobot v3.0 layout and Japanese task-description update
• Benchmark: LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning
• LIBERO authors: Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, Peter Stone
• License: the source repository does not state a license, so none is inferred here

Scale and layout

• Split: train
• Episodes: 1,693
• Frame rows: 273,465
• Task indexes: 40
• Sampling rate: 10 Hz
• Source numeric layout: 377 Parquet shards, 20,264,464 bytes
• TsFile layout: one data/ft_llm_2026_ramen_libero.tsfile table containing all episodes, 16,128,325 bytes
• Table name: ft_llm_2026_ramen_libero

The source meta/tasks.parquet assigns the same general Japanese robot-task sentence to all 40 task indexes. The indexes are retained without inventing more specific task labels.

TsFile schema

Role	Columns	Type and meaning
TIME	Time	INT64 milliseconds, round(timestamp * 1000)
TAG	episode_index, task_index	Original LeRobot episode and task dimensions
FIELD	frame_index, sample_index	Original frame index and source index
FIELD	observation_state_0 ... observation_state_7	Eight FLOAT Panda state values
FIELD	action_0 ... action_6	Seven FLOAT action values

Rows are ordered by episode_index, task_index, and Time. Time starts at 0 for every episode and advances by 100 ms at 10 Hz. Source timestamp is not stored separately because it is exactly recoverable as Time / 1000. Source index is named sample_index; frame_index is retained unchanged.

Videos

Videos are not included here. The original repository contains 3,386 MP4 files, 1,693 for each 256 x 256 AV1 stream:

• observation.images.image
• observation.images.image2

The source video path is videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4. Use episode_index and frame_index together with meta/episodes/chunk-000/file-000.parquet to align TsFile rows to video frames. Hugging Face reports the original videos/ tree as 1.92 GB.

Read example

from tsfile import TsFileReader

reader = TsFileReader("data/ft_llm_2026_ramen_libero.tsfile")
table_name = "ft_llm_2026_ramen_libero"
print(reader.get_all_table_schemas()[table_name])
with reader.query_table(
    table_name,
    ["episode_index", "task_index", "frame_index", "action_0", "observation_state_0"],
    batch_size=65536,
) as result:
    print(result.read_arrow_batch().to_pandas().head())
reader.close()

Citation

@inproceedings{liu2023libero,
  title={LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning},
  author={Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and Liu, Qiang and Zhu, Yuke and Stone, Peter},
  booktitle={Advances in Neural Information Processing Systems},
  year={2023}
}