OriginFlow-AI/origindata-preview-sample

OriginData (preview sample)

OriginData is the world's first large-scale, real-world dataset combining hand pose and force annotations. Spanning 28 domains and 1,170 real-world tasks, it captures how human hands interact with the physical world, providing force, pose, and semantic annotations supported by high-precision multimodal calibration for embodied AI.

This preview contains 100.62 hours across 7,717 episodes, delivered in LeRobot v3.0 format with stereo RGB video, hand pose, head trajectories, finger-force estimates, and hierarchical language annotations.

Key Highlights

• Real-world diversity. Real people performing complete manipulation workflows across everyday life, commercial services, crafts, and industrial assembly.
• Force. The OriginData EGO+EMG pipeline fuses egocentric binocular vision, 16-channel surface EMG, and IMU to derive hand pose and per-finger force, with cross-modal validation.
• High-precision annotations. The capture and fusion pipeline reports synchronization within 3 ms, joint-angle MAE of 3.37°, finger-joint distance error of 4.2 mm, and force magnitude error within 20%.

Video previews

Wrapping Plates and Food with Plastic Wrap Residential Living · Home	Adding Promotion Tags to Instant Noodle Shelves Retail · Supermarket	Packing Drinks for Pickup Food & Beverage · Café	Organizing Sanitized Tableware Food & Beverage · Cafeteria	Buttoning a Shirt Residential Living · Home
Organizing Markers Education · School	Organizing Yogurt Shelves Retail · Supermarket	Organizing Takeout Bags Food & Beverage · Café	Organizing Jump Ropes Sports · Gym	Packing Oden Retail · Convenience Store
Preparing a Drink Food & Beverage · Café	Packing a Parcel for Shipping Manufacturing · Packing Workshop	Organizing Miscellaneous Items Education · School	Applying Promotional Labels Retail · Supermarket	Processing Fiber-Optic Components Manufacturing · Workshop
Organizing Documents Office · Workplace	Frothing Milk Food & Beverage · Café	Assembling a Circuit Board Manufacturing · Electronics Assembly	Demonstrating Infant Massage Techniques Personal Services · Home Care	Repairing a Mobile Phone Personal Services · Electronics Repair

Preview Overview

Attribute	Value
Duration	100.62 hours
Episodes	7,717
Frames	10,867,077
Frame rate	Approximately 30 FPS
Task descriptions	3,323
Format	LeRobot v3.0 with multimodal extensions

Modalities

Modality	Description	Key fields	Documentation
Stereo video	Left and right RGB streams; MP4 / H.264, nominally 1920 × 1080 per view.	observation.images.head_left observation.images.head_right	Video details
Hand pose	Each hand includes 21 3D joint positions in the supplier camera frame (meters), corresponding 2D image coordinates (pixels), and wrist rotations. Joint ordering follows the supplier convention; use the per-hand detection flag to select valid poses.	observation.hand.{side}.joints_3d_camera_m observation.hand.{side}.joints_2d_pixel observation.hand.{side}.wrist_rotation_camera observation.hand.{side}.detected	Pose details
Head trajectory / SLAM	Head/camera position and orientation from the supplied camera trajectory, expressed in the supplier's local SLAM world frame. Position is measured in meters; orientation uses quaternions ordered x, y, z, w.	observation.head.position_in_slam_world observation.head.orientation_in_slam_world_xyzw observation.slam.frame_id	SLAM details
Finger-force annotations	Model-derived force estimates from Fusion2Force, with five scalar channels per hand ordered thumb, index, middle, ring, and little finger. Values are in mN; use the per-hand force-validity flag when selecting samples.	left.force_values right.force_values left.force_valid right.force_valid	Force details
Language annotations	Temporally aligned descriptions at episode, subtask, and atomic-action levels. Frame rows carry the applicable subtask and atomic-action entries in language_persistent; episode metadata stores task descriptions and semantic events with timing and parent-child relationships.	task_index language_persistent Episode metadata: tasks semantic_events_json	Semantic details

{side} is left or right. Timestamps, episode and frame indices, and pose and force validity indicators are also provided.

Download

The Dataset Viewer defaults to four representative LeRobot Parquet shards, one from each data chunk. To load every LeRobot data shard with datasets, choose the full_parquet configuration:

from datasets import load_dataset

frames = load_dataset("OriginFlow-AI/origindata-preview-sample", name="full_parquet", split="train")

The dataset is hosted on Hugging Face.

pip install -U huggingface_hub

# Download the complete dataset
hf download "OriginFlow-AI/origindata-preview-sample" --repo-type dataset --local-dir ./origindata-preview-sample

To explore feature definitions, task descriptions, and episode metadata first:

# Download metadata only
hf download "OriginFlow-AI/origindata-preview-sample" --repo-type dataset --include "lerobot/meta/**" --local-dir ./origindata-preview-sample

Data is provided in LeRobot v3.0 format, with Parquet features, MP4 videos, and multimodal extensions.

Analysis

Real-world coverage. OriginData spans 28 domains, 1,170 real-world tasks. The ring below illustrates the preview source collection through 51 individual scene/task entries; segment sizes reflect source duration.

Episode duration. Episodes average 46.94 seconds, with a median of 31.97 seconds and a range of 0.47–518.29 seconds.

Language coverage. The dataset contains 26,117 subtask annotations and 101,445 atomic-action annotations. All episodes include subtask annotations; 7,654 episodes (99.18%) include atomic-action annotations.

Annotation Quality Evaluation

The Chord / lirong annotation quality report (VLM v2.7, 1-second sampling) evaluates visual finger-contact labels across 7,717 episodes and 364,439 original one-second sample frames. Each frame contributes ten finger-level units (five per hand), for 3,644,390 units in total; all sampled frames were matched successfully.

VLM v2.7 judgments serve as a provisional reference. Binary metrics use only the 2,453,015 units where both labels are explicit:

Metric	Result
Comparable coverage	67.31%
Conditional accuracy	63.49%
Precision	82.38%
Recall	64.59%
F1	72.41%
VLM unknown proportion	32.69%

Usage Notes

• This preview provides model-derived finger-force annotations as five scalar channels per hand.
• Some clips contain occlusion, hands leaving the frame, missing poses, scene transitions, or privacy masks. Use detected, force_valid, and *_validation fields to filter samples.
• Poses retain the original camera coordinate conventions and joint ordering.