robomimic Sim Can Noisy TsFile
Apache TsFile conversion of iantc104/robomimic_sim_can_noisy (https://huggingface.co/datasets/iantc104/robomimic_sim_can_noisy), authored by Ian Chuang (iantc104) and licensed under Apache-2.0.
Source
The train split contains 5,000 can episodes, 328,042 rows, sampled at 20 fps. Source frames use data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet. Source videos use videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4 with agentview and robot0_eye_in_hand streams (256x256 RGB AV1, 20 fps, no audio). Videos remain in the original Hugging Face videos tree and are not included here.
TsFile
- File: data/iantc104_robomimic_sim_can_noisy.tsfile
- Table: iantc104_robomimic_sim_can_noisy
- Rows: 328,042
- meta/ contains source metadata and converted info.json only; source Parquet files are not copied there.
Time is round(timestamp * 1000) in INT64 milliseconds and restarts per episode. timestamp is dropped after mapping; index is renamed sample_index; frame_index is retained for video alignment. episode_index and task_index are TsFile TAG/device dimensions.
Vector fields are flattened to FLOAT32 scalar fields: action (7), action.delta (7), action.absolute (7), observation.state (43), and observation.environment_state (14), with dots replaced by underscores.
Encoding and compression: FLOAT/DOUBLE use GORILLA + LZ4; INT32/INT64/Time use TS_2DIFF + LZ4; BOOLEAN uses RLE + LZ4 when present.
Read Example
from tsfile import TsFileReader reader = TsFileReader('data/iantc104_robomimic_sim_can_noisy.tsfile') with reader.query_table('iantc104_robomimic_sim_can_noisy', ['episode_index', 'task_index', 'frame_index', 'sample_index', 'action_0', 'observation_state_0'], batch_size=65536) as result: print(result.read_arrow_batch().to_pandas().head()) reader.close()
- Downloads last month
- 33