Datasets:

Sichang0621
/

abc_sort_legos_v21

	---
	license: apache-2.0
	task_categories:
	- robotics
	tags:
	- LeRobot
	- yam
	- manipulation
	- bimanual
	- imitation-learning
	configs:
	- config_name: default
	data_files: data/*/.parquet
	---

	# abc_sort_legos_v21

	Teleoperation dataset: sort the legos into containers by color

	A LeRobot v2.1 conversion of the `sort_the_legos_into_containers_by_color`
	task from [XDOF/ABC-130k](https://huggingface.co/datasets/XDOF/ABC-130k),
	prepared for fine-tuning [pi0.5](https://github.com/Physical-Intelligence/openpi)
	(`pi05_base`) on the bimanual YAM platform.

	## Dataset summary

	\| Field \| Value \|
	\|-------\|-------\|
	\| Robot \| yam \|
	\| Episodes \| 500 (first 500 of 4,458 train episodes, sorted by episode uuid) \|
	\| Total frames \| 1,156,733 \|
	\| FPS \| 30 Hz \|
	\| Task \| sort the legos into containers by color \|
	\| Format \| LeRobot v2.1 \|

	## Is this 30 fps?

	Yes — `meta/info.json` reports `fps: 30`, and every camera stream is encoded
	at `video.fps: 30`. But note this is a resampled 30 fps, not a native
	recording rate: in the source ABC-130k MCAP files, the action stream runs at
	~200 Hz, state at ~265 Hz, and cameras at 30–60 Hz depending on station type
	(each stream on its own independent clock). The conversion builds a fixed
	30 Hz tick clock over the overlap window of all streams and does **causal
	floor matching** (the latest message at or before each tick) to align
	everything onto one common 30 Hz grid — actions are subsampled ~6.7:1, faster
	cameras are decimated, and no stream runs below 30 Hz so frames are never
	duplicated.

	## Cameras

	\| Name \|
	\|------\|
	\| `head_camera` \|
	\| `left_wrist_camera` \|
	\| `right_wrist_camera` \|

	Video codec: h264, 640×480 (letterboxed, aspect-ratio preserved). Source
	episodes come from two station types — RealSense (mono top camera) and ZED-X
	(stereo top camera, one eye picked deterministically per episode) — both
	handled by the same conversion.

	## 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. Joint values are absolute positions
	in radians, base → wrist. Gripper is the normalized aperture from
	ABC-130k (0 = closed, 1 = open). `action` holds the commanded joint
	positions (source `/{side}-arm-action` + `/{side}-ee-action` topics);
	`observation.state` holds the measured ones (`/{side}-arm-state` +
	`/{side}-ee-state`). Both are absolute, not delta — pi0.5 applies
	`DeltaActions` internally at train time.

	## How this was converted

	Source episodes are MCAP files (`episode.mcap` per episode). Conversion
	script:
	[`convert_abc_mcap_to_lerobot_v21.py`](https://github.com/Avant-US/openpi) —
	`scripts/convert_abc_mcap_to_lerobot_v21.py`. See "Is this 30 fps?" above for
	the resampling method.

	`meta/episode_ids.json` maps each `episode_index` back to its original
	ABC-130k episode uuid for traceability.

	## Usage

	```python
	from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

	ds = LeRobotDataset("Sichang0621/abc_sort_legos_v21")
	print(ds.num_episodes, ds.num_frames, ds[0]["observation.state"].shape)
	```

	## Attribution and license

	This dataset is **derived from
	[XDOF/ABC-130k](https://huggingface.co/datasets/XDOF/ABC-130k)** (Apache-2.0),
	released alongside the ABC project ([abc.bot](https://abc.bot/),
	[code](https://github.com/amazon-far/abc)). All robot trajectories and
	imagery originate from ABC-130k; this repository contributes only the format
	conversion described above. Please cite the ABC project when using this data.

	Note that the upstream ABC-130k dataset is access-gated on the Hub. Licensed
	under Apache-2.0, consistent with the source.

	## License

	Apache 2.0