At the "Physical AI and Embodied Intelligence Robot Innovation Ecosystem Exchange" forum, part of the China International Fair for Trade in Services, Lingyu Technology (Beijing) Co., Ltd. General Manager and CTO Zhang Jianing delivered a keynote address on September 11, 2026, in Beijing.

Zhang opened by introducing his company's focus on data production and infrastructure for embodied intelligence, emphasizing the collection of high-quality physical interaction data spanning from human behavior capture to real-world robotic operations. He noted that while current mainstream technical approaches are parameter-driven, data serves as the fundamental cornerstone of this field.

The sector faces significant opportunities alongside considerable hurdles, with data scale emerging as the primary challenge. Zhang drew a contrast with large language models, which achieved breakthroughs thanks to massive text datasets built from centuries of human literature and continuous internet content generation. The embodied intelligence field, however, still lacks behavioral data of humans performing tasks in genuine physical environments.

Whether examining home cooking, household chores, or factory assembly line operations, no historical records were ever systematically collected for such activities, creating a severe shortage of training data for embodied models. Currently, most companies train their models on datasets ranging from tens of thousands to hundreds of thousands of hours, with only a select few reaching the latter figure. The industry has set a collective target of accumulating one million hours of high-quality data, and while certain enterprises have proposed ambitious ten-million-hour goals, substantial distance remains before those targets become reality.

The difficulty in acquiring embodied intelligence data stems fundamentally from its nature—unlike internet text that emerges organically through daily activities, this data cannot be generated as a byproduct of other tasks. The industry has yet to identify methods for producing embodied data incidentally while completing unrelated work, making dedicated collection processes essential.

Zhang described the industry's "data pyramid" framework for understanding embodied data structure. The base layer consists of existing 2D videos, which are vast in volume but unsuitable for robot training due to their lack of relevance to human manipulation actions and inconsistent camera angles, rendering them low-quality for this purpose. Moving upward, structured human behavior data relies on UMI, Ego, and similar systems equipped with various peripheral devices that enable simultaneous work and data collection. While operators can gather data during task completion, limitations persist in annotation precision and data modality diversity.

The next tier comprises simulation data, which offers the advantage of generating diverse scenarios within virtual environments, though bridging the gap between simulation and reality demands substantial research investment. At the pyramid's apex sits real-machine data, offering the highest quality but suffering from critically low production efficiency. Zhang summarized this landscape as the current state of embodied intelligence data, identifying the acquisition of high-quality real-world data as the industry's key challenge for the coming phase.

Lingyu Technology's work directly addresses these pain points through a unified technical foundation supporting two parallel data tracks designed to solve embodied intelligence data acquisition. The first track, the Ego series, focuses on lightweight human behavior collection in real-world settings, with particular emphasis on hand manipulation—the most critical capability for achieving dexterous robotic operations. The company has developed precise annotation tools tracking 6DoF positioning across 21 key hand points, while simultaneously recording force and tactile data during grasping, holding, retrieving, and placing actions. This effort has produced hardware including smart gloves and headbands, with the product line also expanding toward whole-body intelligence to comprehensively capture human movement.

The second track, Sonic2, establishes a virtual-real interaction foundation connecting physical robots with simulation environments. The remote teleoperation solution enables operators to control physical robots from distant locations while robots execute tasks in real environments and synchronously capture data. This system relies on proprietary StarTraq® positioning technology, achieving millimeter-level whole-body accuracy and supported by over 280 global technical patents.

Highlighting the Ego product portfolio, Zhang introduced the NOLO EgoCapture smart headband, equipped with five cameras—four fisheye RGB units and one front-facing ToF sensor. The device captures environmental information in real time while recording data, with algorithms performing automated annotation. Four solutions address hand and full-body data labeling: the bare-hand tracking solution elevates 6DoF positioning accuracy for 21 key hand points to millimeter precision; the OmniGlove smart glove integrates StarTraq® positioning for full-hand millimeter accuracy while embedding piezoresistive tactile sensor arrays across palms and fingers to capture pressure distribution throughout hand movements; the OmniBand external tracking system pairs IR arrays with smart headband cameras, while remaining compatible with third-party smart gloves, exoskeletons, and other hand-tracking devices; and the OmniMotion full-body motion capture solution deploys 11 tracking nodes with the smart headband to record complete human body positioning data.

The comprehensive solution offers two core advantages: millimeter-level tracking accuracy across the entire pipeline, and microsecond-level temporal synchronization for multimodal data. These precision levels stem from a decade of technical and product development in spatial computing.

Turning to the second track, Zhang detailed the remote teleoperation scheme built on the NOLO Sonic2 series. While traditional teleoperation requires operators and robots to share physical space—with operators wearing XR devices whose movements drive robot actions—this remote solution liberates the operator from proximity constraints. An operator in Beijing, for example, can control robots deployed in other cities to complete tasks. In commercial scenarios, robots stationed in physical convenience stores independently handle checkout and stocking duties while remote personnel oversee operations, simultaneously collecting authentic scene data during real business activities. Data generated in genuine environments demonstrates significantly improved quality compared to training-ground collections.

The solution maintains approximately 100-millisecond latency across diverse network conditions, fully supporting remote operational control. Binocular anti-distortion processing on robot-returned video feeds provides operators with immersive visual experiences enabling precise task execution. Beyond physical robot teleoperation, the company has also developed simulation-based teleoperation, integrating with NVIDIA Isaac Sim's high-fidelity physics simulation platform to support multi-person whole-body and multi-robot tracking for collaborative simulation scenarios.

Complementing the hardware ecosystem, Lingyu Technology has developed a complete data platform offering end-to-end processing pipelines covering upload, storage, automated annotation, and manual quality inspection. The company is also advancing future-oriented initiatives, including extending robot teleoperation from ground to space. A ground-space collaborative remote expert system is currently operational aboard China's space station, enabling ground-based scientists to guide astronauts through physics and chemistry experiments. The company is now exploring whether ground scientists can remotely operate robots inside the station to independently complete all experimental work, which would dramatically enhance station efficiency. While this ambitious goal demands further optimization of end-to-end latency and other technologies, the strategic value is immense, and Zhang extended an invitation for interested enterprises to collaborate on advancing this frontier.

Zhang concluded his presentation with thanks to the audience.