From Data Collection Needs to Device Design: Launch of Nexdata's Own Ego-Centric Device, SeerFusion One

As VLA models, World Models, and robot learning keep advancing, the need for Ego-Centric Data in Physical AI evolves from first-person videos to multimodal data, which consists of vision, depth, pose, motion, and interactions.

Based on its practical experience from tens of thousands of data collection tasks, Nexdata designed SeerFusion One specifically for large-scale Physical AI data collection. It was made taking into account the specific needs of multimodal data collection, real-time pose tracking, sensor synchronization, data quality control, and cross-device spatiotemporal alignment.

Multimodal Ego-Centric Recording with Real-Time 6DoF Tracking

SeerFusion One incorporates stereo RGB cameras, four mono fisheye cameras, optional ToF depth, and an array microphone to enable simultaneous recording of first-person visual, depth, and audio data.

At the same time, the device incorporates an IMU and onboard SLAM to track 6DoF head pose in real time with online calibration. According to the test results, APE does not exceed 0.3%, whereas ARE does not exceed 0.5°.

By recording vision, depth, and pose at the same time, SeerFusion One does not just record the field of view of the operator, but also their position and movements in the scene. This provides more complete Human Demonstration Data for training VLA models, World Models, and robot learning systems.

Hardware-Level Synchronization and Real-Time Quality Control

Accuracy of temporal alignment between different modalities of Physical AI data directly impacts further data processing and model training.

SeerFusion One relies on hardware-level timestamps to ensure millisecond-level synchronization of the data collected by onboard sensors, including video, depth, IMU, and pose.

The device also allows real-time QC for such problems as frame drops, pose loss, tracking out-of-range events, and timestamp mismatches. Online calibration helps reduce the impact of temperature fluctuations, vibrations, and mechanical displacement during long data collection sessions.

By performing part of the quality control process during data collection, issues can be detected and corrected at an earlier stage, reducing the amount of invalid data and repeated collection.

From Head-Mounted View to Hands and Whole-Body Motion

Physical AI data does not just consist of first-person view. To train models for complex manipulation tasks, it is also necessary to understand spatial relations between the head, body, hands, tools, and surrounding objects.

SeerFusion One can act as the central node of a multi-device recording system and be combined with wrist cameras, gloves, 6DoF trackers, and UMI grippers to extend first-person data to hand motion, body pose, tool trajectories, and end-effector actions.

Using hybrid spatial tracking, poses of different devices can be mapped into a common coordinate system and a common time reference to establish the spatial relations between the head, body, hands, and tools.

In the case of head-hand coordination, time synchronization between different devices can be less than 1 ms.

Thanks to this approach, multi-device recording is no longer limited to independent data streams, but can create multimodal Human Demonstration Data with consistent spatiotemporal relations.

Designed for Physical AI Data Collection

The development of SeerFusion One is based on Nexdata's practical experience in data collection.

From multimodal recording and real-time 6DoF tracking to hardware-level synchronization, real-time QC, online calibration, and cross-device spatiotemporal alignment, the key features of the device are oriented toward one purpose—making Ego-Centric data collection more reliable, efficient, and scalable for Physical AI data production.

With the launch of SeerFusion One, Nexdata has expanded its Physical AI data capabilities further upstream—from data collection, annotation, and model validation to self-developed data collection devices.