Datasets:
SO101 Pick Place TsFile
Apache TsFile edition of jonhpark/so101_pick_place, a LeRobot v2.1 SO101 manipulation dataset. The numeric trajectory table contains robot actions, states, timing, frame positions, and source episode/task identifiers.
Source and attribution
- Original repository owner: jonhpark
- License: Apache-2.0
- Robot type:
so101; LeRobot codebase version:v2.1 - Split:
train; 50 episodes; 29,792 frame rows; one task; 30 FPS; one data chunk - Task
0: Grasp a skyblue doll and put it in the white bowl. - The source card does not provide a homepage, paper, or completed BibTeX citation.
Data layout
The table is jonhpark_so101_pick_place with 29,792 rows and 50 episode/task devices. Source frame Parquet follows data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet; the TsFile path is data/jonhpark_so101_pick_place.tsfile.
| Column | Role | Type | Meaning |
|---|---|---|---|
Time |
TIME | INT64 milliseconds | round(timestamp * 1000), restarting at zero for each episode |
episode_index |
TAG | STRING | Source episode identifier |
task_index |
TAG | STRING | Source task identifier |
frame_index |
FIELD | INT64 | Frame position within the episode |
sample_index |
FIELD | INT64 | Source index, renamed for the scalar table |
action_0 ... action_5 |
FIELD | FLOAT | Flattened source action[6] |
observation_state_0 ... observation_state_5 |
FIELD | FLOAT | Flattened source observation.state[6] |
The six action dimensions are main_shoulder_pan, main_shoulder_lift, main_elbow_flex, main_wrist_flex, main_wrist_roll, and main_gripper; the state vector uses the same order. The source timestamp is omitted because it is represented by Time / 1000 seconds. No trajectory rows, episodes, tasks, action dimensions, or state dimensions are removed.
Videos and alignment
Camera data is not included in this TsFile repository. The original Hugging Face dataset keeps 100 MP4 streams under:
The source path template is videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4. Use episode_index and frame_index to align each numeric row with its front or top 30 FPS video.
Read example
from tsfile import TsFileReader
reader = TsFileReader("data/jonhpark_so101_pick_place.tsfile")
with reader.query_table(
"jonhpark_so101_pick_place",
["episode_index", "task_index", "frame_index", "sample_index", "action_0", "observation_state_0"],
batch_size=1024,
) as result:
batch = result.read_arrow_batch()
print(batch.to_pandas().head())
reader.close()
- Downloads last month
- 44