Real-World Data Collection
Multimodal, diverse, and high-quality data collection tailored for generative foundation models.
SERVICES
End-to-end, high-quality data services for foundation models, embodied AI, and world models — collected in the real world, verified and annotated by domain experts, delivered ready for large-model training.
COLLECT · ANNOTATE · EVALUATE
Domain-Expert Annotation
Specialized human-in-the-loop labeling, RLHF, and custom data curation.
End-to-End Evaluation
Robust benchmarking, bias mitigation, and model safety testing.
VARIOUS PARADIGMS · SCALE TO GROUND TRUTH
Embodied data has no internet-scale corpus to scrape — it has to be manufactured, one interaction at a time, through the right collection paradigms and rigorous quality control.
B.1 / PROBLEM
Language models can learn from internet text, and vision models can learn from vast amounts of images; but robots lack a ready-made "physical world interaction database" that can be directly scraped. Thus, embodied AI data must be collected one by one in the real world.
Embodied data collection isn't a single method — it's a combination of different paradigms, each suited to a different collection scenario and transfer requirement.
Ego-Centric Capture
First-person data captured via wearable devices/cameras, recording human actions and interactions in real environments — scalable without requiring a physical robot.
- HOME VISUAL + ACTION
- 4M+ units / 40K+ hours
- TEAM
- 550+ person team
- WEARABLE DAILY-WORK
- 7,000+ hours
UMI Gripper Capture
Captures human–object manipulation trajectories via a UMI (Universal Manipulation Interface) gripper — combining human dexterity with a robot-executable action representation.
EGO-CENTRIC UMI-GRIPPER TELEOPERATION
- ON-TIME DELIVERY
- 100%
Real-Robot Teleoperation
Direct teleoperation of physical robot hardware, capturing action trajectories that transfer straight to deployment — no sim-to-real gap.
CAPABILITY
Supports real-robot teleoperation across multiple robot platforms and gripper types, with task scenarios, robot platforms, operating tasks, and data modalities configurable per project — covering the full execution chain from equipment deployment and operator training, through task design and scaled collection, to data QC and delivery.
Dexterous Hand
Fine-manipulation data for multi-fingered dexterous hands — covering grasping, in-hand rotation, and bimanual coordination — filling the coverage gap that gripper-only data leaves in fine-motor scenarios.
MODALITIES
Captures multimodal data spanning finger-joint states, fingertip poses, hand poses, object poses, and vision — extensible to tactile, force/torque, and contact-state data depending on hardware capability, covering tasks from basic grasping through high-difficulty dexterous manipulation.
Simulation
Real-world 3D assets for robotics simulation, covering everyday objects, tools, furniture, and environmental objects to provide diverse and realistic assets for simulation environments.
DELIVERY
Supports both direct selection from an existing library of real-world 3D assets and custom 3D digitization of specified objects. Assets can be customized to project requirements for object category, scale, and quality, then delivered in formats ready for integration into simulation platforms.
Collected. Covered. That's the easy part.
Calibrated and verified — that's what makes it count.
01 / SCALE
The bottleneck stopped being the model a while ago — it's whether there's enough usable data. Embodiment-bound collection hits a hard ceiling on hardware costs and fixed workstations; OrphLux's no-embodiment collection network runs 100K+ hours and 200+ contributors in parallel on a single case, unconstrained by hardware.
- HOURS / MONTH · NO-EMBODIMENT COLLECTION
- 100K+
- DATA COLLECTORS COLLECTING DAILY
- 10,000+
- REAL-WORLD DATA COLLECTION SITES
- 100+
02 / DIVERSITY
Diversity isn't about stacking up more scenes — it's about covering every scene type that real life and real work actually run on, with each task captured as something genuinely happening in the moment instead of staged actions.
- LOGISTICS & WAREHOUSING
- RETAIL
- INDUSTRIAL PRODUCTION
- FOOD & HOSPITALITY
- HOME LIFE
- OFFICE
- TRAINING / ELDER CARE
Down to concrete sites such as actual residences, offices, supermarket backrooms, sorting centers and warehouses, factory floors, restaurant/hotel kitchens, training facilities, elder-care settings, and nail/hair salons.
Every task captured is something "genuinely happening in that setting, with a clear operational purpose" — sorting packages, boxing/unboxing, picking and shelving in logistics; food prep and service in food & hospitality; material handling and basic assembly on a production line — all drawn from the actual workflows of real practitioners on the job.
03 / GROUND TRUTH
Raw video isn't training data by itself — actions, poses, contact points, and success labels have to be correct, or the model learns the noise instead of the pattern.
- TEMPORAL ANNOTATION ACCURACY
- 98%+
- MULTIMODAL ANNOTATION ACCURACY
- 95%+
- TIMESTAMP ALIGNMENT ERROR
- ≤0.5s
04 / CUSTOM HARDWARE
Because different data-collection tasks require different hardware and final data outputs, we help customers select, configure, and customize hardware based on their task scenarios, target data modalities, and delivery requirements, and further process raw collected data into the structured data they require.
- HARDWARE SELECTION
- Identify suitable ego devices, UMI devices, sensors, and other devices based on project requirements.
- HARDWARE CUSTOMIZATION
- Customize camera count, viewpoints, mounting structures, sensor combinations, and other hardware configurations.
- DATA POST-PROCESSING & CUSTOMIZATION
- Convert raw video, sensor signals, and device state data into customer-defined structured data, such as trajectories, poses, action states, event segments, contact information, and task-level labels.