Datasets:

Sichang0621
/

8ml_vial_place_30fps_fixed

This is a corrected fork of ttotmoon/8ml_vial_place_30fps. The source dataset was recorded with a YAML camera-config bug that assigned the wrong RealSense serial number to each wrist camera, so the videos under observation.images.left_wrist_camera/ actually contained physical-right-wrist footage (and vice versa). This fork swaps them back so labels match physical reality. See § What was fixed below.

Dataset summary

Field	Value
Robot	yam
Episodes	197
Total frames	229,860
FPS	30 Hz
Task	place the vial into the stand
Format	LeRobot v3.0

Cameras

Name
`head_camera`
`left_wrist_camera`
`right_wrist_camera`

Video codec: hevc.

State space (`observation.state`, shape `[14]`)

Index	Name
0	`left_joint_0`
1	`left_joint_1`
2	`left_joint_2`
3	`left_joint_3`
4	`left_joint_4`
5	`left_joint_5`
6	`left_gripper`
7	`right_joint_0`
8	`right_joint_1`
9	`right_joint_2`
10	`right_joint_3`
11	`right_joint_4`
12	`right_joint_5`
13	`right_gripper`

Action space (`action`, shape `[14]`)

Index	Name
0	`left_joint_0`
1	`left_joint_1`
2	`left_joint_2`
3	`left_joint_3`
4	`left_joint_4`
5	`left_joint_5`
6	`left_gripper`
7	`right_joint_0`
8	`right_joint_1`
9	`right_joint_2`
10	`right_joint_3`
11	`right_joint_4`
12	`right_joint_5`
13	`right_gripper`

State and action are 1:1 index-aligned.

Note: Gripper dimensions may use different units between state (raw motor radians) and action (teleoperator command space). The robot's JointMapper rescales between these at command time. Left and right grippers may also have different value ranges if the physical hardware travel differs.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("<repo_id>")
print(ds.num_episodes, ds.num_frames, ds[0]["observation.state"].shape)

What was fixed

The original dataset's recording-time YAML mis-assigned wrist-camera serial numbers. As a result, in the source dataset:

videos/observation.images.left_wrist_camera/ contained physical-right-wrist footage
videos/observation.images.right_wrist_camera/ contained physical-left-wrist footage

This fork applies a one-shot swap so the labels match the physical camera mounts.

What was touched

Video directories renamed: videos/observation.images.left_wrist_camera/ ↔ videos/observation.images.right_wrist_camera/
meta/episodes/chunk-000/file-000.parquet — swapped the videos/observation.images.{left,right}_wrist_camera/{chunk_index, file_index, from_timestamp, to_timestamp} column groups (4 cols × 2 cams) and the stats/observation.images.{left,right}_wrist_camera/ {min, max, mean, std, count} column groups (5 cols × 2 cams).

What was NOT touched

data/*.parquet carries only state, action, episode_index, frame_index, timestamp, index, task_index — no per-camera columns.
meta/info.json — both wrist features have identical schemas, so the JSON is byte-identical to the source.
meta/stats.json, meta/tasks.parquet — no per-camera content.
observation.state / action columns — left_joint_* / right_joint_* refer to arm identity from the CAN bus, independent of camera labeling. The arms themselves were always correctly identified; only the wrist cameras' labels were swapped.

Downstream impact

Any checkpoint trained on the source (e.g. ttotmoon/yam-vial-place-pi05-v1) learned the buggy wrist assignment and will not produce correct behavior when served with a physically-correct camera-config YAML. Retrain SFT from pi05_base on this dataset for a consistent serving setup.