On August 18, Li Yi, founder and CEO of 51WORLD, officially launched AperData, an integrated hardware‑software embodied AI data infrastructure, at the "Physical AI Blueprint 2030" launch event, with an initial price of RMB 5,100 per set.

As a subsidiary jointly incubated by 51WORLD and Union Image, AperData is committed to building a neutral, open, integrated "collection terminal + data governance engine + open data interface" that connects the entire pipeline from real‑world data collection to processing, quality inspection, computation, evaluation, and dataset delivery – transforming data production from one‑off collection projects into standardized, traceable, and scalable infrastructure.

Reconstructing Embodied AI Data Collection – A 10x Leap in Quality and Efficiency

As more embodied AI systems move from laboratories to the real world, the industry faces a very real problem: where do robots get enough high‑quality data that can actually be used for training, to ensure they accurately understand the physical world and make reliable autonomous decisions?

The scale, quality, and collection efficiency of data have become the core bottlenecks in embodied AI evolution!

Traditional hardware‑software separated collection approaches suffer from a natural disconnect between data quality and production efficiency. Hardware devices, collection software, and backend data processing are handled by different systems, making it difficult to achieve unified standards for time synchronization, spatial calibration, and data formats across devices. Problems often only surface during data upload or even the training phase, leading to rework and repeated collection of invalid data. At the same time, collection, quality inspection, upload, and processing rely on manual handoffs, creating fragmented workflows that cannot support large‑scale, high‑efficiency data production.

What AperData reconstructs is not a single collection device, but the very way effective data is produced for embodied intelligence: the integrated hardware‑software solution ensures 99% physical trajectory consistency while significantly reducing invalid collection and rework, delivering a 10x leap in unit effective data cost and production efficiency compared to traditional teleoperation.

High‑Precision Collection Hardware: Ensuring Data Quality at the Source

In addition to the AperEgo head‑worn series, the AperData hardware products launched this time also include the optional AperWristCAM wrist camera and the AperFinger gripper collection handle, catering to diverse data collection needs across different scenarios.

As the core terminal for first‑person perspective data collection, AperEgo uses multi‑camera arrays working in coordination with IMUs to build more comprehensive environmental perception capabilities. Taking the AperEgo 4‑camera model as an example: it uses four global‑shutter cameras to create a first‑person panoramic collection capability, with a total horizontal field of view exceeding 270°, vertical field of view exceeding 155°, and downward field of view exceeding 100°, enabling complete recording of the collector's surrounding environment. The front binocular cameras use a 65mm baseline, and through high‑precision mounting and calibration design, errors are controlled to within 0.1mm*, meeting the spatial consistency requirements for depth perception and VLA training.

AperEgo achieves full‑sensor hardware synchronization with synchronization accuracy of <200μs, integrating quad‑camera vision, IMU at >200/500Hz, and multi‑channel audio into synchronized collection. With microsecond‑level hardware timestamps and high‑concurrency buffering, it ensures temporal consistency of multimodal data during long‑duration continuous collection.

AperWristCAM further records "how human hands act." By adding close‑up hand perspective through the wrist camera, it forms a full‑spectrum "human eye + human hand" collection combination with AperEgo, extending from environmental understanding to hand movements and object interaction – providing more complete data inputs for robot motion replication, VLA training, and multimodal dataset production.

The AperData hardware product matrix, with its multi‑perspective, multimodal, high‑synchronization hardware architecture, ensures that both "seeing" and "doing" in the real world can be fully recorded, meeting the diverse data needs of embodied intelligence hardware, dexterous hands, VLA/world model training, and other fields.

AperOS: Full Lifecycle Data Closed Loop

The core problem that the AperOS data platform solves is establishing a standardized closed‑loop process that connects project management, task assignment, on‑site collection, data upload, preprocessing, quality inspection and annotation, computation, evaluation, and dataset delivery – forming a closed loop covering the entire data production lifecycle. Through cloud‑edge collaboration, the collection side handles device connection, task execution, and on‑site inspection, while the cloud handles data governance and subsequent processing, reducing data transfer and manual handoffs between different systems and turning collected data into true data assets.

Traditional data collection often follows a "collect first, process later" model: data is uploaded to the cloud after collection, and only then do quality checks and data cleaning reveal which data is actually valid. Invalid segments – where actions weren't completed, the collection process was interrupted, or the collection hadn't formally started – are also uploaded, only to be identified and discarded during post‑processing. Once problems are found, not only does re‑collection become necessary, but additional network transmission, storage, and cloud processing costs are incurred.

AperOS moves this process forward to the collection side. Beyond real‑time checks on quality metrics such as clarity, exposure, frame drop rate, and multi‑channel synchronization, it can also perform pre‑processing and data cleaning during collection, identifying and discarding invalid collection segments while retaining truly valuable data. For formal collections with issues, they can be detected and re‑collected on the spot in real time. If a 30‑second collection only has 10 seconds of truly valid content, AperOS can filter out the invalid portions directly on the edge device before uploading the valid data. This means data is not "collected first and cleaned later" – invalid data is reduced at the source from entering the subsequent pipeline.

This edge‑side processing fundamentally changes the cost structure of data production: reducing invalid data uploads lowers network bandwidth and cloud storage pressure; reducing the volume of data entering the cloud also means a corresponding decrease in the computing power required for subsequent data computation and processing. On this foundation, AperOS continuously transforms cleaned valid data into trainable data assets through unified data ingestion, time alignment, standardization, high‑precision computation, and full‑chain traceability – ensuring that every collection, every transmission, and every unit of computing power is spent on truly valuable data.

Toward 2030: From Infrastructure to Global Network

The launch event also unveiled AperData's five‑year product capability roadmap toward 2030.

At the launch event, AperData also unveiled its five‑year product capability roadmap toward 2030. Over the next five years, AperData will continue to evolve both software and hardware in tandem, steadily improving the foundational capabilities for embodied AI data production. In 2026, the software side will establish foundational data annotation and quality inspection toolchains, while the hardware side will complete the first‑generation mass production of head‑worn collection kits. In 2027, the software will evolve toward a multimodal data fusion platform covering vision, force, and posture, while the hardware will expand to full‑body multimodal data collection suits. In 2028, the software will further develop AI‑powered automated annotation and scenario generation engines, and the hardware will move toward lightweight, wireless collection devices. In 2029, the software will build simulation‑augmented data closed loops (Real2Sim2Real), and the hardware will launch high‑precision dexterous hand collection terminals. By 2030, AperData will further evolve into fully automated data factories, embodied foundation model data engines, open data ecosystem APIs, and end‑to‑end quality assessment systems, while also building low‑cost terminals, globally distributed collection networks, multi‑form‑factor hardware adaptation kits, and large‑scale mass‑production supply chains – transforming from a data production system into an open data infrastructure for embodied intelligence, making the real world a continuously renewable data source for robot learning and evolution.

*Data based on internal test estimates; actual figures subject to mass‑production verification.

Product Trial / Purchase

Scan the QR code below.

About AperData

AperData (Shenzhen AllScene Intelligence Technology Co., Ltd.) is a joint venture incubated by 51WORLD and Union Image, dedicated to building a neutral, open, integrated hardware‑software data infrastructure for embodied intelligence – making the real world the data foundation for embodied AI.

ROBOTECH 2026 Intelligent Robotics Expo will be held from October 27–29, 2026 at the Shenzhen World Exhibition & Convention Center (Bao'an) .

Union Image will make a major appearance at the show. Scan the QR code below to register as a visitor now – see you in Shenzhen this October for a face‑to‑face conversation with Union Image!