LingYu-Robotics/lingyu-realworld-robotics-dataset

LingYu Real-World Robotics Dataset

1. Introduction

LingYu Real-World Robotics Dataset provides real-world teleoperation data for research in imitation learning and task-specific policy adaptation. In this first release, we are open-sourcing 139.4 hours of high-quality bimanual first-person teleoperation data (13,200+ episodes) spanning 15 tabletop task families and 22 subsets, covering deformable-object manipulation, fine-grained and precision manipulation, as well as object sorting and placement.

In addition, we plan to make a further dataset of the real-world retail data available to the research community through an application-based access program with manual review. The data was collected in an active supermarket during regular operations. Access will be limited to approved research projects and governed by responsible-use and privacy-preserving requirements. Our longer-term goal is to expand deployment in real-world settings and collect diverse operational demonstrations under appropriate privacy safeguards. This dataset is not disclosed in this release; please find more access information in Section 4.2.

We open-source a series of datasets with the intention of supporting research on policy training, fine-tuning, and adaptation to real-world environments. If you're working on an interesting research project that could benefit from our open-source dataset, we'd love to hear from you. We plan to provide up to 1,000 hours of data free of charge to selected projects. Contact us at contact@dexteleop.com for further discussion.

2. Dataset at a Glance

Item	Value
Total duration	139.4 h
Episodes	13,263
Subsets	22, each a self-contained LeRobot dataset in its own folder
Repository	LingYu-Robotics/lingyu-realworld-robotics-dataset
Total size	~9.16 TB (tabletop, 39,789 MP4 + 13,263 Parquet)
Robot	LingYu TA2 bimanual robot: 2 × 7-DoF arms with parallel grippers
Collection	VR teleoperation
Cameras	3 stereo RGB views: head-mounted stereo camera (3840 × 1920) and left / right wrist stereo cameras (2560 × 800)
Control / recording rate	45 Hz, all streams synchronized to the same clock
Proprioception	Joint position, velocity, effort; end-effector pose (both arms) — 62 dims
Actions	Joint position & velocity commands, gripper command, end-effector target pose — 62 dims
Format	LeRobot dataset format v2.1 (Parquet + MP4/HEVC)
License	CC BY-NC-SA 4.0

3. Research Value

3.1 Key Facts about this Dataset

• Rich, synchronized modalities: Each episode contains three high-resolution RGB video streams together with full bimanual proprioception (joint position, velocity, and effort for 14 arm joints and 2 grippers), end-effector poses of both arms, and the corresponding teleoperation commands (joint position and joint velocity targets, gripper commands, and end-effector target poses), all resampled to a common 45 Hz timeline.

• First-person VR View: Teleoperation is performed through a VR headset using the robot’s onboard camera feeds, giving the operator the same first-person visual perspective available to the robot. No third-person views or external scene information are introduced. This preserves the robot’s deployment-time observation space.

• Diverse and long-horizon tasks: The release spans 15 tabletop task families covering deformable-object manipulation, fine-grained and precision manipulation, and object sorting and placement. Episode lengths range from a few seconds for single pick-and-place actions to several minutes for long-horizon sorting and stacking tasks.

• Quality control: Episodes are screened by human review, and automatically rejected on arm-command jumps, command/state tracking errors, interrupted or restarted recordings, and incomplete sensor streams. Episodes are released whole; abnormal ones are removed rather than trimmed.

3.2 Real-World Data from Regular Operation

In addition to the tabletop datasets included in this release, we have accumulated teleoperation data from routine food-sampling operations in a commercial supermarket. These data extend the collection beyond controlled laboratory or purpose-built environments to active retail workflows.

• In-the-wild data collection: Captures robot manipulation under the variability and operational constraints of an active retail environment.

• Task-grounded demonstrations: Links manipulation behaviors to concrete retail procedures, supporting policy learning for structured real-world workflows.

We plan to make a further dataset of these real-world retail data available to the research community through an application-based access program with manual review. Access is subject to responsible-use and privacy-preserving data-use requirements.

4. Task Coverage

4.1 Tabletop Tasks

15 task families released as 22 standalone subsets (Jenga is split by tower height, and the shape-insertion subset contains two instruction variants).

Subset	Category	Task instruction	Episodes	Hours
fold_towel_into_box	Deformable	Fold the white towel and place it in the box.	2,714	25.5
fold_airplane_box	Deformable / assembly	Fold the airplane box from flat to fully assembly.	950	24.4
fold_socks	Deformable	Fold the pair of socks in front of you and place the folded pairs together in the container.	309	1.7
put_tennis_ball_in_bucket	Placement	Put the tennis ball into the bucket.	2,289	19.4
sort_blocks_by_print	Sorting	Sort tabletop blocks into bins by print: numbers & math symbols → pink, English letters → blue, unprinted → green.	337	12.0
insert_shape_blocks	Fine-grained	Insert all the colored shape blocks into their matching slots on the puzzle board (tilted / vertical pick-and-place, 2 variants).	963	11.0
sort_screws_and_nuts	Fine-grained / sorting	Sort the screws and nuts into two different cups.	457	3.9
place_medal_in_box	Fine-grained	Place the medal in the box and close the lid.	453	3.4
sort_plush_toys_into_bins	Sorting / selection	Sort the plush toys into their designated bins, leaving non-plush objects untouched.	406	1.2
sort_desk_items	Sorting	Place the stationery into the container and put the crumpled paper into the desktop trash bin.	195	1.1
place_fruits_on_plates	Placement	Take one apple, one orange, and one pear from the baskets and place each on its corresponding plate.	377	1.9
place_orange_on_plate	Placement	Pick up an orange from the table and place it on the plate.	618	1.0
place_apple_on_plate	Placement	Pick up an apple from the table and place it on the plate.	234	0.4
place_white_pen_in_holder	Fine-grained / selection	Pick up the white-bodied pen and place it into the pen holder, leaving all black-bodied pens untouched.	249	0.7
stack_jenga_tower_2layers	Fine-grained, long-horizon	Stack all the Jenga blocks on the table into a 2-layer tower, with at most 3 blocks per layer.	134	0.3
stack_jenga_tower_3layers	Fine-grained, long-horizon	… into a 3-layer tower …	219	0.9
stack_jenga_tower_4layers	Fine-grained, long-horizon	… into a 4-layer tower …	101	1.1
stack_jenga_tower_5layers	Fine-grained, long-horizon	… into a 5-layer tower …	310	2.7
stack_jenga_tower_6layers	Fine-grained, long-horizon	… into a 6-layer tower …	401	3.9
stack_jenga_tower_7layers	Fine-grained, long-horizon	… into a 7-layer tower …	420	6.5
stack_jenga_tower_8layers	Fine-grained, long-horizon	… into an 8-layer tower …	525	7.0
stack_jenga_tower_9layers	Fine-grained, long-horizon	… into a 9-layer tower …	602	9.4
Total			13,263	139.4

The Jenga subsets use the full instruction: "Stack all the Jenga blocks on the table into an N-layer tower, with at most 3 blocks per layer."

4.2 Real-World Retail Environment Data

In addition to the dataset presented in this work, we have collected a separate set of teleoperation demonstrations during routine food-sampling activities at a commercial supermarket. Operators remotely controlled the robot to deliver food samples to customers during regular store operations. These demonstrations were recorded using the same robot, cameras, and data schema as the tabletop subsets.

This separate collection of real-world retail data is outside the scope of the dataset presented here. We plan to make it available to approved research projects through an application-based access program with manual review, subject to responsible-use and privacy-preserving requirements.

Access description

Researchers interested in accessing the data may apply by providing a brief description of their research objectives, intended use, project or institutional context, and relevant data-handling practices. Requests will be reviewed individually based on the relevance of the proposed research and the applicant’s ability to comply with the applicable data-use and confidentiality requirements.

The underlying retail site is intentionally not identified in this release. Approved users must not attempt to identify or disclose the site, individuals, or other confidential information, and may use the data only in accordance with the applicable terms. If your research could benefit from access to this type of real-world retail data, please contact us at contact@dexteleop.com.

5. Data Description, Data Format & Schema

5.1 Directory layout

All subsets are published in one dataset repository, each as a self-contained LeRobot v2.1 dataset in its own top-level folder. The folder name is identical to the subset name used throughout this document:

LingYu-Robotics/lingyu-realworld-robotics-dataset/
├── README.md
├── fold_towel_into_box/          # one folder per subset, named exactly as the subset
│   ├── meta/
│   │   ├── info.json             # features, shapes, fps, totals, path templates
│   │   ├── episodes.jsonl        # one line per episode: episode_index, tasks, length
│   │   ├── episodes_stats.jsonl  # per-episode feature statistics (min / max / mean / std / count)
│   │   └── tasks.jsonl           # task_index -> natural-language instruction
│   ├── data/
│   │   └── chunk-000/
│   │       └── episode_000000.parquet   # one Parquet file per episode, one row per frame
│   └── videos/
│       └── chunk-000/
│           ├── observation.images.head_camera/episode_000000.mp4
│           ├── observation.images.left_color/episode_000000.mp4
│           └── observation.images.right_color/episode_000000.mp4
├── place_apple_on_plate/
│   └── ...                       # same structure
└── ...

Episodes inside a subset are grouped into chunks of 1,000 (chunk-000, chunk-001, ...). episode_index and the global index are local to each subset — two subsets do not share an index space.

Note on the Hugging Face dataset viewer: the repository is a collection of standalone LeRobot datasets rather than a single flat Parquet table, so the automatic dataset viewer may not render it. Download the subset folder you need and load it with LeRobot as shown in section 6.

5.2 Features

Feature	Type	Shape	Description
observation.state	float32	(62,)	Measured robot state, see 5.3
action	float32	(62,)	Teleoperation command, same layout and names as observation.state, see 5.3
observation.images.head_camera	video (HEVC, yuvj420p)	1920 × 3840 × 3	Head camera
observation.images.left_color	video (HEVC, yuvj420p)	800 × 2560 × 3	Left camera
observation.images.right_color	video (HEVC, yuvj420p)	800 × 2560 × 3	Right camera
timestamp	float32	(1,)	Time since episode start (s), frame_index / 45
frame_index	int64	(1,)	Frame index within the episode
episode_index	int64	(1,)	Episode index within the subset
index	int64	(1,)	Global frame index within the subset
task_index	int64	(1,)	Index into meta/tasks.jsonl

5.3 State / action layout (62 dims, identical across all subsets)

observation.state and action share the same 62-dim layout and the same dimension names. Dimensions that carry no information for a given stream are filled with 0 (marked below). All values are in SI units: radians, radians per second, newton-metres and metres.

Index	Names	observation.state (measured)	action (command)
0–6	left_joint{1..7}_position	Left arm joint positions (rad)	Left arm joint position targets (rad)
7	left_gripper_position	Left gripper motor position (rad)	0
8–14	right_joint{1..7}_position	Right arm joint positions (rad)	Right arm joint position targets (rad)
15	right_gripper_position	Right gripper motor position (rad)	0
16–22	left_joint{1..7}_velocity	Left arm joint velocities (rad/s)	Left arm joint velocity commands (rad/s)
23	left_gripper_velocity	Left gripper velocity (rad/s)	0
24–30	right_joint{1..7}_velocity	Right arm joint velocities (rad/s)	Right arm joint velocity commands (rad/s)
31	right_gripper_velocity	Right gripper velocity (rad/s)	0
32–38	left_joint{1..7}_effort	Left arm joint efforts (Nm)	0 (velocity control)
39	left_gripper_effort	Left gripper effort (Nm)	Left gripper command (Nm), see 5.4
40–46	right_joint{1..7}_effort	Right arm joint efforts (Nm)	0 (velocity control)
47	right_gripper_effort	Right gripper effort (Nm)	Right gripper command (Nm), see 5.4
48–50	left_ee_position_{x,y,z}	Left end-effector position (m)	Left end-effector target position (m)
51–54	left_ee_orientation_{x,y,z,w}	Left end-effector orientation (unit quaternion)	Left end-effector target orientation
55–57	right_ee_position_{x,y,z}	Right end-effector position (m)	Right end-effector target position (m)
58–61	right_ee_orientation_{x,y,z,w}	Right end-effector orientation (unit quaternion)	Right end-effector target orientation

Notes:

• Control mode. The arms are velocity-controlled. Both the joint position targets (0–14) and the joint velocity commands (16–30) are recorded and meaningful; joint effort is not commanded, so the corresponding action dims are 0.

• End-effector frame. End-effector poses are expressed in the shoulder frame of the respective arm. Orientations are unit quaternions in (x, y, z, w) order.

• Gripper. The gripper is commanded through the effort channel (action[39], action[47]); action[7] and action[15] are 0. Use observation.state[7] / [15] for the measured gripper state: like the arm joints, this is a motor position in radians, not a normalised opening width.

• Fixed layout. All subsets share this exact 62-dim layout, so any number of subsets can be concatenated or jointly trained without adapters.

5.4 Gripper command normalization

The gripper command is stored as effort in Nm (nominally in [-1.6, +2.0]). It can be converted to a normalized trigger value in [0, 1] with the exact inverse of the platform's trigger→effort curve:

import numpy as np

def gripper_effort_to_trigger(effort: np.ndarray) -> np.ndarray:
    """Convert gripper effort command (Nm) to the normalized [0, 1] trigger value.

    Exact inverse of the platform's trigger->effort curve (v2, both arms).
    Deliberately unclipped so the effort->trigger->effort round trip is exact;
    efforts slightly outside [-1.6, +2.0] map slightly outside [0, 1].
    """
    return np.where(effort > 0, 0.10 * (1.0 - effort / 2.0), 0.10 - effort * 0.90 / 1.6)

5.5 Data quality

Episodes are screened by human review before release. Recordings with abnormal length, arm-command jumps, command/state tracking errors, interrupted or restarted takes, or incomplete sensor streams are removed. Every episode is released in full — no frames are trimmed or interpolated.

5.6 Repository and folder naming convention

All subsets live in one dataset repository, LingYu-Robotics/lingyu-realworld-robotics-dataset, and each subset is a top-level folder inside it. The folder name is identical to the subset name used throughout this document:

LingYu-Robotics/lingyu-realworld-robotics-dataset/<subset>

<subset> follows a fixed verb_object[_modifier] rule in snake_case:

Prefix	Meaning	Examples
fold_	deformable object folding / assembly	fold_towel_into_box, fold_socks, fold_airplane_box
stack_	object stacking	stack_jenga_tower_4layers
insert_	insert into a receiving structure	insert_shape_blocks
sort_	multi-way sorting into targets	sort_screws_and_nuts, sort_blocks_by_print
place_	single pick-and-place / placement	place_apple_on_plate, place_medal_in_box
put_	placement into a container	put_tennis_ball_in_bucket

Rules for future releases:

1. New task → new top-level folder with the same verb_object[_modifier] rule. No date, version or batch number in the name.

2. More data for an existing task → same folder. Append episodes to that subset and bump the repository revision; user code keeps working unchanged.

3. A different robot or a different environment adds a scope prefix, so subsets stay distinguishable. The 62-dim schema and LeRobot v2.1 format stay identical across scopes.

4. Subsets that differ only by a discrete parameter use a _<parameter> suffix rather than separate names (stack_jenga_tower_2layers … 9layers), so they sort and group naturally.

Because all subsets sit in one repository, keep the folder list flat — no nested per-task grouping. A flat list keeps the folder name as the single key for loading a subset, and both LeRobot and plain hf download address a subset with one path.

6. Quick Start / Download & Use Instructions

6.1 Download

pip install -U "huggingface_hub[cli]"

# whole collection (all subsets, ~9.16 TB)
hf download LingYu-Robotics/lingyu-realworld-robotics-dataset --repo-type dataset --local-dir ./lingyu

# only the subsets you need
hf download LingYu-Robotics/lingyu-realworld-robotics-dataset --repo-type dataset --local-dir ./lingyu \
  --include "fold_towel_into_box/*" "place_apple_on_plate/*"

# a single task family, e.g. all eight Jenga subsets
hf download LingYu-Robotics/lingyu-realworld-robotics-dataset --repo-type dataset --local-dir ./lingyu \
  --include "stack_jenga_tower_*/*"

Each subset is self-contained, so you can also copy ./lingyu/<subset> anywhere and load it locally.

6.2 Load with LeRobot

The data uses LeRobot dataset format v2.1, and each subset folder is a complete dataset root. Use a LeRobot release that reads v2.1 datasets, or convert with LeRobot's v2.1 → v3.0 conversion script for newer releases.

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset  # v2.1-era import path

# `root` must point at the subset folder; metadata is read from <root>/meta and data from <root>/data
ds = LeRobotDataset(repo_id="LingYu-Robotics/lingyu-realworld-robotics-dataset",
                    root="./lingyu/fold_towel_into_box")
frame = ds[0]
print(frame["observation.state"].shape, frame["action"].shape)  # torch.Size([62]) torch.Size([62])
print(frame["observation.images.head_camera"].shape)            # (3, 1920, 3840)
print(frame["task"])

repo_id is only used as an identifier (and as a fallback source when root has no meta/), so with the subset downloaded locally this loads fully offline. Passing only repo_id without root will not work, because the repository root has no meta/ — it is a collection of subsets. In LeRobot 0.1.x the dataset class reads <root>/meta/info.json, <root>/data and <root>/videos directly, so any subset folder is a valid dataset root on its own.

Because every subset shares the same 62-dim layout, subsets can be concatenated for multi-task training without any feature remapping.

6.3 Read the raw files

import json
import numpy as np
import pyarrow.parquet as pq

root = "./lingyu/fold_towel_into_box"
info = json.load(open(f"{root}/meta/info.json"))
names = info["features"]["action"]["names"]

table = pq.read_table(f"{root}/data/chunk-000/episode_000000.parquet")
action = np.stack(table["action"].to_numpy(zero_copy_only=False))            # (T, 62)
state = np.stack(table["observation.state"].to_numpy(zero_copy_only=False))   # (T, 62)

joint_pos_cmd = action[:, list(range(0, 7)) + list(range(8, 15))]             # (T, 14)
gripper_trigger = gripper_effort_to_trigger(action[:, [39, 47]])             # (T, 2), see 5.4

7. Intended Use & License

Intended use. This dataset is intended for research on robot learning, including imitation learning, policy pre-training, fine-tuning, and adaptation to real-world environments, as well as benchmarking of manipulation policies.

Out-of-scope use. Use for identifying or tracking individuals, or any use that violates applicable laws or the license terms.

License. This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

8. Citation & Contact

If you find this dataset useful in your research, please cite:

@misc{lingyu_realworld_robotics_2026,
  title        = {LingYu Real-World Robotics Dataset},
  author       = {{LingYu Robotics}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/LingYu-Robotics}}
}

For dataset-related questions and further academic discussions, please contact contact@dexteleop.com.

For business inquiries, please contact business@dexteleop.com.