Case Study: Nexdata UMI Data Collection

Nexdata recently completed a UMI-based data collection project for Physical AI model development. The project helped the client obtain human operation data from real home environments to support robot algorithm optimization and improve task execution capabilities in household scenarios.

Project Description

The project was designed around natural household task flows. Unlike isolated or repetitive single-action data, the collection process was organized using natural human operation logic and required collectors to complete specific tasks.

The project was carried out in different home spaces, including bedrooms, kitchens, balconies, living rooms, entryways, and bathrooms. Tasks were organized according to room sequence and household activity logic, allowing the collected data to reflect natural operation processes in home environments.

Examples of the tasks performed by collectors include:

Folding clothes

Organizing kitchen counters

Tidying sofas

Arranging household items

Completing room-based cleaning and organization tasks

To increase scenario diversity, the project was carried out in different apartments with small, medium, and large household environments. Each household environment supported up to 25 hours of data collection, and the first batch is planned to cover more than 500 household scenarios.

Collection Requirements

During project execution, Nexdata focused on natural task logic and data continuity.

Collectors were required to perform tasks in a purposeful and sequential way according to realistic household activity paths, such as moving from the bedroom to the kitchen, balcony, living room, entryway, and bathroom depending on the task design.

This collection strategy captured not only manipulation actions, but also room transitions, object placement differences, spatial layout variations, and continuous task execution processes. These aspects are important for training Physical AI models to understand household environments and perform long-horizon tasks.

Project Outcomes

Nexdata completed full-scenario household robot task data collection, covering more than six types of home spaces, such as kitchens, living rooms, bedrooms, balconies, entryways, and bathrooms.

The collected data met the client’s requirements and provided comprehensive support for robot algorithm optimization, task execution improvement, and Physical AI model development.

Multi-room, multi-task, and goal-oriented operation data helped the client build a reliable data foundation for household robot learning and scenario adaptation.

Project Capacity

The first stage of the project is planned at a scale of 20,000 hours of valid data, with collection currently ongoing.

After the household scenario collection is completed, the project is expected to expand into other commercial and service scenarios, providing a broader data foundation for developing Physical AI models in different environments.

Nexdata’s UMI Data Collection Services for Physical AI

Nexdata provides customized UMI data collection for Physical AI model development, covering various scenarios, tasks, spatial layouts, and collection requirements.

Based on structured project design, scalable collection capacity, and quality control processes, Nexdata helps robotics teams obtain high-quality data for robot learning, manipulation model training, VLA/VLM development, humanoid robot applications, and household task execution.