LuteCorp/soft-yubi-100h

Soft Yubi 100h

Project page: https://lute-corp.github.io/soft-yubi-site/

Bimanual manipulation demonstrations collected with handheld YUBI grippers (UMI-style capture): a demonstrator wears one gripper per hand, each tracked by a Meta Quest, with a wrist camera on each gripper and a fixed top camera. Every episode is stored as end-effector state and action at the tool tip, so the same action space can drive a robot carrying the YUBI gripper.

The data comes in two splits, each one LeRobot v3.0 dataset per task:

• mid-train: 50 pick-and-place tasks, for large-scale mid-training.
• sft: 4 tasks for supervised fine-tuning.

Dataset Statistics

Cameras	top + left wrist + right wrist
Video codec	AV1 (libsvtav1), yuv420p
Resolution	640x480, all cameras
Frame rate	30 fps
Action space	20-dim: [left_xyz(3), left_rot6d(6), left_gripper(1), right_xyz(3), right_rot6d(6), right_gripper(1)]
Total hours	mid-train: 92.03 h · sft: 7.22 h
Trajectories	mid-train: 47,555 episodes · sft: 3,563 episodes

Dataset Structure

soft-yubi-100h/
  mid-train/<task>/    # one LeRobot v3.0 dataset per task
  sft/<task>/
    meta/
      info.json        # features, fps, shapes
      stats.json       # normalization stats (mean, std, min, max, q01..q99)
      tasks.parquet    # the task prompt (= the task name)
      episodes/        # per-episode index ranges and stats
    data/chunk-*/file-*.parquet
    videos/observation.images.<camera>/chunk-*/file-*.mp4

Each frame carries:

Feature	Shape	Description
observation.state	(20,)	per hand: tool-tip xyz (m), 6D rotation, gripper opening normalized to [0, 1] per gripper from its encoder calibration (0 = closed, 1 = fully open); left then right
action	(20,)	the next frame's observation.state (the last frame repeats)
observation.images.top	480x640x3	fixed top camera
observation.images.wrist_left	480x640x3	left gripper camera
observation.images.wrist_right	480x640x3	right gripper camera
is_fresh.*	(1,)	whether each stream was newly measured at this frame or held from the previous one

State names: {left,right}_{x,y,z,r1..r6,gripper}. Poses are in the Quest tracking frame. Each pose is the Quest-tracked hand_root transformed to the original YUBI end effector by TIP_XYZ = (0.0784, 0, -0.0004) m and TIP_ROT = rpy(0, π/2, 0).

Dataset Breakdown

Split	Task	Episodes	Hours	Blacked Top Cam
mid-train	almonds	940	2.06	partial (619 / 940 eps)
mid-train	baby_wipes	789	1.50	partial (249 / 789 eps)
mid-train	batteries	588	1.79	true
mid-train	bonitki_coconut_cookies	1,758	2.90	false
mid-train	bottle_oshee	482	1.35	true
mid-train	bottle_rivella	830	1.91	true
mid-train	bottle_water	866	1.92	true
mid-train	candy_mamba	1,371	1.72	false
mid-train	cereal_cheerios	903	1.91	false
mid-train	cereal_chocapic	1,313	1.94	false
mid-train	cereal_corn_flakes	970	1.88	false
mid-train	cookies_hit	1,033	1.93	false
mid-train	crackers	1,195	1.95	false
mid-train	dish_scrub	1,007	1.95	false
mid-train	dish_soap	1,025	1.63	false
mid-train	dish_sponge	850	1.94	partial (401 / 850 eps)
mid-train	dried_mango	1,030	1.94	partial (208 / 1030 eps)
mid-train	ear_buds	715	1.79	partial (490 / 715 eps)
mid-train	film_roll	1,407	1.91	false
mid-train	fruit_puree	1,436	1.93	false
mid-train	glue	1,068	2.00	false
mid-train	jasmine_rice	1,153	1.94	false
mid-train	kicthen_roll_velvet	1,187	1.89	false
mid-train	light_bulb_black	847	1.68	true
mid-train	light_bulb_white	596	1.81	true
mid-train	measuring_tape	805	1.94	true
mid-train	milk_tube	1,035	1.95	false
mid-train	mirror_clean_spray	1,248	2.20	true
mid-train	muesli_bar	1,714	1.93	false
mid-train	pads	684	1.72	true
mid-train	pasta_chifferi	556	1.35	false
mid-train	pasta_farfalle	985	1.90	true
mid-train	pasta_tagliatelle	1,120	1.89	true
mid-train	peas	1,215	1.96	false
mid-train	pencils	850	1.92	partial (650 / 850 eps)
mid-train	power_tool_battery	473	1.63	true
mid-train	ramen	1,067	1.89	false
mid-train	sealant_remover	803	1.76	true
mid-train	socket	665	1.88	true
mid-train	tea_green	742	1.97	true
mid-train	tea_red	1,119	1.94	false
mid-train	tile_spacers	1,073	1.88	true
mid-train	tissue_box	798	1.65	true
mid-train	tissue_packet	803	1.97	true
mid-train	toilet_paper_queen	16	0.04	false
mid-train	toothpaste	575	1.69	true
mid-train	trash_bags	1,055	2.01	partial (571 / 1055 eps)
mid-train	wall_paint_brush	1,000	1.90	false
mid-train	wd40	890	1.99	partial (458 / 890 eps)
mid-train	wrench	905	1.89	partial (426 / 905 eps)
mid-train total	50 tasks	47,555	92.03
sft	box	704	1.93	false
sft	porridge	631	1.39	false
sft	wire	999	1.58	false
sft	zupa	1,229	2.32	false
sft total	4 tasks	3,563	7.22

Blacked Top Cam: true if every episode of the task has an all-black top camera, partial if some recording days do.

Loading the Dataset

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(repo_id="mid-train/ear_buds", root="soft-yubi-100h/mid-train/ear_buds")
frame = dataset[0]
frame["observation.state"], frame["action"], frame["observation.images.top"], frame["task"]

Tested with lerobot==0.5.1.

License

This dataset is released under the PolyForm Noncommercial License 1.0.0.