Data Services
Scalable production of high-quality embodied AI training data
TA2 Real Robot Data Training Center
Real-world data collection center providing real-robot data at scale for embodied AI model training
LingYu TA robots and DexUMI are deployed in real physical environments to carry out data collection at scale, supplying real-robot data for model training. The plan is to accumulate a million hours of real-robot data and provide research institutions and model developers with a standardized physical testing environment.
Data Quality Assurance
Time Alignment
Sub-millisecond
Sub-millisecond synchronization across all sensors, with nanosecond-level triggering for 6 cameras; camera-exposure-to-memory latency as low as 40 ms
Spatial Alignment
0.1 mm
Repeatability of 0.1 mm and absolute positioning accuracy of 1 mm, supporting cross-body data alignment and reuse
Force Control Collection
0–10 N
Supports 0–10 N force-controlled collection and high-frequency compliant torque control, recording contact force information
Information Density
Multimodal
Combines force control data, 4K head binocular, 2K wrist binocular and remote-operation eye tracking data
LingYu Data Platform
Data management platform covering collection, upload, review, annotation and training preparation.
LingYu Data Platform covers the full robot-data workflow, from device preparation, task assignment, data collection and upload through to review, annotation and training preparation. It provides device operation monitoring, collection efficiency analysis, data asset statistics, annotation task scheduling, team management and performance analysis, and supports ROS Bag as well as custom formats that have been adapted, helping teams manage data production in one place.
Core Capabilities
Full-process Data Management
Core business chain: device preparation → create/assign collection tasks → robot collection → data upload → data warehousing → data review → annotation tasks → annotation submission → review approval → training data preparation
Data Overview
Presents core metrics such as data collected, tasks executed, device coverage and production efficiency. Supports filtering by time range, device and operation, with multi-dimensional rankings and trend analysis.
Review Workbench
Filter tasks by collection time, operation, device, operator and data status. Reviewers can view left-wrist, head and right-wrist video channels along with sensor data, and submit review conclusions against data quality standards.
Data Asset Overview
View the scale and status flow of data assets accumulated on the platform from a management perspective. Data status funnel: collected → uploaded → review passed → annotated → training-ready. Each stage shows data duration, share of collected data, reduction versus the previous stage, and stage conversion rate.
Device Cluster Management
View device count, status distribution, basic information and real-time operating metrics. Device states include working, idle, faulty, stopped, error and offline; real-time metrics include CPU, temperature, memory and battery.
Collection Efficiency Center
Tracks participating operators, task count, data volume, data duration, estimated effective duration and effective duration ratio. Supports rankings by operator and hourly distribution of working time.
Annotation Task Hall
Annotators can claim tasks in the task hall and enter an annotation page centered on video and a timeline. They can complete segment annotation, confirm, modify or delete AI pre-annotation results, create new annotations, save progress and submit for review.
Task & Team Management
Administrators can create, assign, edit and track annotation tasks, maintain team groupings of annotators and reviewers, and organize task allocation and performance review by team.
Performance Overview
Management can review overall completion of annotation tasks and individual performance. Supports daily, weekly, monthly, yearly, all-time and custom time ranges, with task creation trends, status distribution, average quality score and average utilization.
Server Storage Monitoring
Monitor total storage, used space, usage ratio and recent usage trends to help assess storage capacity and expansion needs.
Customized End-to-End Data Service
End-to-end service from requirements analysis to data delivery, covering real-robot data collection that integrates mobility and manipulation
Robot versions, sensor combinations and work environments are configured to the customer's scenario, supporting multi-robot parallel collection, sub-millisecond hardware synchronization, cross-region remote operation and professional operating demonstrations. After collection, the data goes through quality verification, format standardization, AI-based cleaning and intelligent annotation, and is finally delivered as a standardized data package that can be fed into the customer's training pipeline.
Hardware: Precision, Stability and Synchronization
Custom Automotive-grade Cameras
Built with an Active Alignment (AA) process to improve the stability of intrinsic and extrinsic camera parameters and reduce calibration drift over long-term use.
Global Active Sync Triggering
Active synchronized triggering yields accurate exposure timestamps, reducing time-synchronization error while the arms are in motion.
Low-latency High-precision Motion Control
A self-developed motion control algorithm balances positioning accuracy against response speed, improving the motion fidelity of trajectory data.
Robot Calibration and Frame Alignment
Robot calibration and coordinate frame alignment are performed uniformly, giving the algorithm side standardized intrinsic/extrinsic data and improving consistency and reusability across devices.
Data: Real, Long-horizon and Diverse
Long-horizon Tasks in Real Scenarios
Covers continuous long-horizon tasks in real scenarios, supporting data collection for complex tasks.
Chained Nested Task Design
A chained, nested task design covers the composite operating workflows found in real applications.
Diversity and Quality Control
Diversity requirements are set across scenarios, objects, actions and boundary conditions, with an independent data quality review standard that improves the completeness, consistency and training usability of delivered data.
7-Step Loop
Define Task Objectives
Define scenarios, task types and data specifications, and confirm the collection environment, target data volume, modality combinations and quality standards.
Build Collection Plan
Configure robot versions, sensor combinations and work environments to the requirements, and define collection workflows and staffing.
Scalable Data Collection
Supports multi-robot parallel collection with sub-millisecond hardware synchronization, plus cross-region remote operation and professional operating demonstrations. Daily collection exceeds 1,000 hours and 80,000 episodes.
Quality Control
Verification against 1 mm absolute accuracy automatically screens out anomalous data; collection progress and device status are monitored in real time.
Format Standardization
Standardized MCAP output with timestamp alignment and metadata annotation, compatible with ROS Bag.
Fine-grained Annotation
Supports AI-based cleaning and intelligent annotation, combined with timeline slicing and a multi-level label system.
Delivery & Integration
Delivers a standardized data package that can be fed into the customer's training pipeline, together with a data quality report and usage documentation.
Contact Us
Product inquiries, data services, partnerships and press all start here. Tell us what you need in the form.
Your details are used for this inquiry only.