Tactile Module
Records high-precision, distributed tactile interaction data.
PaXini Omnimodal Embodied Intelligence Dataset
OmniSharingDB
OmniSharing DB is a omni-modal embodied dataset covering tactile, visual, audio, text, and joint data, providing comprehensive support for embodied AI models from physical perception to semantic cognition.China’s first government-supported, fully compliant embodied AI cross-border data project, strictly adhering to the Cybersecurity Law, Data Security Law, and PIPL.
Legally Exportable
Strictly compliant with China’s Cybersecurity Law, Data Security Law, and Personal Information Protection Law (PIPL), with official approval for cross-border data transfer.
High-speed Transfer
PaXini collaborates with leading global cloud service partners to build a “fast lane” for large-scale data delivery.
Compliance Channel
Establishes a compliant pipeline from collection, processing, and certification to cross-border transfer, enabling global circulation for embodied data.
Dataset Features
Comprehensive data diversity for embodied AI development
Joint Module
Records joint angles from the collection glove and tracks global trajectory data.
Visual Module
Provides synchronized multi-view, high-frame-rate video streams that comprehensively capture the task scenario.
Spatial Trajectory
Process multi-view synchronized video streams and annotate corresponding spatial trajectories using estimation algorithms.
Voice Module
Voice data collected before task execution, describing the task.
Text Module
Structured textual information obtained by converting the collected Voice data.
Collection Devices
PXCap V Ultra
Proprietary multi-modal hand sensing system integrates sensing glove, camera array and IPC to acquire hand kinematics, full-domain tactile data and task videos.
82 DoFFull-hand tactile coverageUltra-compact anti-magnetic encoder
PXCap III
PXCap III combines isomorphic data collection and execution with high-precision tactile and motion capture. It provides reliable, high-quality data for embodied AI in real-world scenarios.
Isomorphic Collection and ExecutionHigh-Precision SensingStrong Actuation PerformanceHuman-Centric
Collection Solutions
In-Domain Collection
Targeting dexterous manipulation scenarios, human-centric high-precision sampling generates high-fidelity haptic data to close model data loops.
Operation-focusedHuman-CentricPrecision-guaranteed
In the Wild Collection
Open collection builds upon In the Wild data collection, relaxes scenario constraints to enrich the model’s environmental awareness, improves long-sequence task planning performance, and evolves iteratively to expand diversified data collection solutions.
Egocentric DataLow-Cost Large-Scale CollectionReal-Environment Collection
Scene Types
15+N
Task Categories
100+
Task Categories
1000+
All Scenes
Office
Home
Healthcare
Medical
Rehabilitation
Logistics
Education
Entertainment
3C Manufacturing
Restaurant
Agriculture
Business
Retail
Disaster Rescue
Automotive
Industrial
Meta Action
Data Generalization
Cross-embodiment capabilities and comprehensive data characteristics
Data Features
Simulation Env
High-Fidelity physical sensing&Sim Platform
Digital Assets
High Fidelity 3D Gasuusian splatting
Multi-Adaptability
High-Accuracy Retargeting&Response
Different Method
Comparing different data collection approaches for embodied AI
approach
Efficiency
Motion Fidelity
Tactile Fidelity
Cross-embodiment Naturalness
Human-Centric Data Collection
(PaXini PMEC Hyper-collection Technology)
Teleoperation
(Traditional Technology)
Simulated Data Synthesis
(Traditional Technology)
Data Format
Hierarchical dataset organization across all processing stages
DF-1
The overall input: raw data after preprocessing and quality inspection
DF-2
1st output: DF-1 with encoder and tactile data parsed; adds bimanual and object poses; includes both action and observation
The observation layer in DF-2 and DF-2R is the state data in the episode, while action is one frame behind observation. To ensure both arrays have equal length, the last frame in action is repeated.
DF-2R
2nd output: DF-2 retargeted to a dexterous hand model
The observation layer in DF-2 and DF-2R is the state data in the episode, while action is one frame behind observation. To ensure both arrays have equal length, the last frame in action is repeated.
DF-3
3rd output: converts DF-2R to the LeRobot dataset format; can be used for VLA model training
Dataset Structure DF-1
▼
task_list▼
task_XXX▼
episode_0_HHMMSS_CollectionDate_RoomID_CollectorsID.hdf5▼
dataset-
metaTask Description▼
observation-
audioCompressed Audio Stream, Including Text Description▼
state▼
left_hand▼
encoder-
data-
timestamp▼
tactile-
data-
timestamp▼
right_hand▼
encoder-
data-
timestamp▼
tactile-
data-
timestamp▼
image-
rgbd_rgb_extrinsic▼
RGB_CameraXXX-
timestamp-
data-
extrinsics-
intrinsics▼
RGBD_CameraXXX-
timestamp-
extrinsics-
inner_extrinsics▼
color-
data-
intrinsics▼
depth-
data▼
left-
data-
intrinsics▼
right-
data-
intrinsics-
episode_1_HHMMSS_CollectionDate_RoomID_CollectorsID.hdf5-
episode_2_HHMMSS_CollectionDate_RoomID_CollectorsID.hdf5/
......▼
task_XXX/
......Omnisharing Toolkit
Comparing different data collection approaches for embodied AI
PX Pose
the pose estimation and processing module
Raw visual data parsing
DF-1 format parses RGBD camera information and aligns multi camera temporal data to provide input for pose estimation.
Pose Estimation of Wristbands and Objects
Binocular infrared depth estimation, wristband automatically detects pose, and objects need to be manually labeled with masks to estimate pose.
PX Post-Process
Parses and converts after all raw data is obtained
Parsing
Processes the raw data output from PX Pose and generates the DF-2 data format.
Retargeting
Retargets DF-2 data to the embodiment and outputs DF-2R.
Conversion
Processes DF-2R data and exports it in the LeRobot data format.
Data Closed Loop
Cross-embodiment capabilities and comprehensive data characteristics
Real & Simulated Data Verification
Publications
Cross-embodiment capabilities and comprehensive data characteristics
PaperDatasetCode
OmniVTLA: Vision-Tactile-Language-Action Model with Semantic-Aligned Tactile Sensing
Paper
TacCompress: A Benchmark for Multi-Point Tactile Data Compression in Dexterous Manipulation
PaXini · Shenzhen
+86 755 2357 4593
PaXini · Shanghai
+86 21 5456 1536
PaXini · Tianjin
+86 22 8241 1882
mkt@paxini.com China
sales_global@paxini.com Overseas
hr@paxini.com