Real-World Multimodal Data Collection Service
Collected at real work sites, then processed, annotated and quality-checked. We deliver data ready to train your robots.
From collection to delivery, in one seamless flow.
Our data infrastructure covers device management, on-device quality checks, data transfer, annotation, QA, data augmentation and visualization, turning recordings of real work into multimodal embodied data ready for model training.
First-person cameras, third-person cameras and data collection gloves record visual, motion, spatial and contact information during real tasks.
Broad, deep, accurate and stable
Covers real tasksHandles different operators, objects, tools and environments, including one-handed, two-handed and continuous tasks.
MultimodalRGB, depth, IMU, hand joints, 6-axis force, tactile data and task semantics are recorded in sync.
Reconstructs space and physicsReconstructs the spatial relationships and physical states of hands, objects and environment, and outputs accurate camera poses, point clouds and force data.
Standardized productionA reproducible pipeline from task definition and collection validation to processing, annotation, QA and version control.
Natural work, recorded as it happens at real sites
Real environmentsCollected in real work environments, preserving natural work rhythm, environmental noise and task context.
Complex manipulationSupports two-handed coordination, tool use, fine contact and continuous tasks.
Wide variationCovers variation in operators, objects, spatial layouts, motion paths and outcomes.
Large-scale managementManaged by site, date, task and batch, with continuous tracking of volume, quality and progress.
We support a wide range of sites, including food service, agriculture, manufacturing and services.
What each data record contains
What's Inside
From a single task, we record the following four types of data in sync and deliver them together.
Video & depthFirst- and third-person RGB video, plus raw and optimized depth and IMU data.
Hands & force3D position and orientation of 21 hand joints, plus 6-axis force and tactile data capturing contact and force changes.
Position & shapeCamera poses, 6-DoF poses of manipulated objects and a point cloud of the whole scene.
SemanticsTask instructions, subtasks, atomic actions, states and outcomes, annotated per time segment.
Whether data is usable for training depends on spatial, motion and force accuracy
- Optimized depth and 3D space : reconstructs scene depth and geometry, plus the 3D layout of hands, objects and surroundings
- Accurate hand pose : records hand position, orientation and continuous motion trajectories in 3D space
- 6-axis force and tactile : captures contact and force changes during tasks with 3-axis forces and 3-axis moments
From collection to delivery
Define task scope, site conditions, data items and acceptance criteria
Validate equipment setup, procedures and output format with a small pilot
Scale up collection with unified SOPs and parameters
Time sync, depth optimization, pose computation, task segmentation and semantic annotation
Check for anomalies and deliver with batch, version and quality records
Service menu
Dedicated datasetsWe design data around your model goals and specific tasks and build a dedicated dataset.
Custom collectionWe plan and run collection at scale to match your data schema, equipment setup and acceptance criteria.
Processing & annotationWe apply time sync, depth processing, pose computation and annotation to your existing data.
Ongoing additional collectionBased on training and evaluation results, we fill data gaps such as new tasks, hard cases and failure examples.
Documentation