Nexdata’s Integrated Physical AI Data Solution: Hardware, Platforms and Services

Data collection and processing are often bottlenecks for Physical AI teams.

Even with the right hardware, capturing every part of an interaction is not easy. Multiple sensor streams need to be accurately aligned with each other. Then, once the initial data batches have been processed and annotated, teams need to maintain that quality as data collection scales up.

Complete capture, accurate alignment, reliable annotation and consistent delivery are vital when creating training datasets.

With 15 years of experience in data collection and annotation, Nexdata has served more than 1,000 clients. It combines collection hardware, data management, annotation and professional services in one integrated solution. Clients can choose individual parts of this solution or opt for the whole package to create their own Physical AI data pipelines.

Configure Collection Hardware to Suit the Task

Grasping requires not only locating an object but also approaching, touching and moving it. Nexdata provides data collection setups for both robot teleoperation and human demonstrations, with visual, motion and tactile data capture depending on the task.

For robot teleoperation, Nexdata offers dexterous hand data collection kits featuring Linker Hand O6, L6, L20 and L20 Lite models and the Linker Arm A7 robotic arm. Available equipment includes teleoperation arms, force-feedback gloves and inertial motion capture gloves. Orbbec Gemini 335L overhead cameras and Gemini 2 wrist cameras capture visual observations, while the setup also records joint states, tactile signals and torque data, depending on the configuration.

For human demonstrations, Nexdata uses its proprietary egocentric device with six built-in cameras. Two RGB cameras forming a stereo pair, with a resolution of 1920 × 1280 and horizontal and vertical fields of view of 157° and 102°, respectively, work together with four fisheye cameras to capture close-range manipulation and the surrounding environment.

The device supports depth, IMU, audio and pose data capture. Its proprietary visual SLAM (VSLAM) system provides real-time 6DoF pose estimation and spatial localization. Millimeter-level spatial calibration and sub-millisecond temporal calibration support alignment across sensors. Motion capture gloves and wrist cameras can be added to record hand movements and additional views.

Adapt Collection Environments and Services to the Task

The collection environment is as important as the hardware itself. Object placement, workspace layout and interaction conditions affect the actions performed and captured.

For teleoperation tasks, Nexdata’s 10,000-square-meter data collection facility supports more than 20 scenarios, including warehouses, supermarkets and pharmacies. The team configures hardware, objects and tasks according to project requirements and carries out collection and quality checks in controlled environments.

For egocentric and UMI tasks, Nexdata organizes collection in real-world settings, capturing actual layouts, object arrangements and interaction conditions. Clients specify their target scenarios and tasks, and Nexdata coordinates the hardware and operators needed to collect relevant demonstrations.

Integrate Collection, Data Management and Annotation Processes

Processing recorded data into training datasets requires close coordination across these steps. Nexdata’s solution connects collection devices with the data management and annotation platforms.

Collection devices integrate directly with the data management platform, allowing data to be uploaded with encryption after collection. After passing quality checks, post-processing and acceptance review, the data is transferred to the annotation platform. Completed annotations are returned to the data management platform, where recordings and annotations are managed together.

The process covers nine steps:

Task creation → Task claiming → Data collection → AI-based quality checks → Human-in-the-loop quality checks → Cloud-based post-processing → AI-based pre-annotation → Human-in-the-loop annotation → Format conversion

Annotation requirements, including task instructions and action steps, can be adjusted to suit model training needs. AI automation and human review support quality control and data processing throughout the workflow. Integrating the devices and platforms reduces manual data transfers and repeated integration work while keeping recordings and annotations centrally managed.

Nexdata provides hardware, platforms and services for device deployment, data collection and data processing. Clients can use the complete solution or choose the parts they need.

Four Categories of Off-the-Shelf Datasets

In addition to Bespoke data collection and annotation services, Nexdata offers four categories of off-the-shelf datasets for training and evaluation:

Teleoperation data: 10,000 Hours Embodied AI Dextrous Manipulation Data-Teleoperation

UMI data: 50,000 Hours UMI Embodied Manipulation Dataset

Egocentric data: 10000-Hours 6-Camera Ego Embodied AI Dataset

Simulation data: 100 Million Sets of 3D Scene & Object Data

Whether you need an off-the-shelf dataset for your experiments, Bespoke data collection, or hardware and platforms for ongoing data collection, Nexdata can help.

Contact Nexdata to discuss your target tasks, collection environments and delivery requirements.