On July 17, the 9th World Artificial Intelligence Conference (WAIC) commenced at the Shanghai World Expo Center. As a pivotal annual event in the AI sector, this year's conference attracted major domestic and international companies specializing in large models, AI applications, and hardware. Leitech (ID: leitech), through its AI-centric new media platform 'Leitech AGI (leikejiagi),' sent a reporting team to Shanghai to provide on-the-ground coverage.

(Image source: On-site photography by Leitech)

At this year's WAIC, embodied AI took center stage. Xiao Lei spent the entire day exploring the exhibition hall, marveling at humanoid robots from over twenty companies capable of performing tasks like pouring coffee, moving boxes, and screwing bolts. Yet, amidst the dazzling array, the robots began to blend together after a while.

One booth, however, stood out: Mifeng Technology's. Unlike others, there were no humanoid robots, no agile robotic hands, nor any moving parts. Instead, on display were a head-mounted data collection device and a handheld gripper-like apparatus. After a brief experience, Xiao Lei realized that Mifeng wasn't showcasing what robots could do, but rather how future robots would learn to perform tasks.

(Image source: On-site photography by Leitech)

Mifeng's booth, though small, was meticulously organized. On the left were the two core hardware components of the MEgo series. In the middle was an interactive experience area, and on the right, a large screen displayed real-time processing from the data governance platform. Xiao Lei first tried the handheld gripper, the MEgo Gripper, with a simple task: pick up the items in front. For this seemingly straightforward action, the gripper's 200° fisheye lens, 3D tactile sensors, and nine-axis IMU worked in unison, recording vision, depth, posture, force, and motion trajectory in real-time. The adjacent screen reconstructed the operation with millimeter-level precision, digitizing every detail, from the force applied by each finger to the arc traced through the air.

(Image source: On-site photography by Leitech)

Next up was the MEgo View head-mounted device: five cameras on the head provided a 300° panoramic view, while a wrist-mounted camera focused on hand details, with all seven channels capturing data simultaneously. After a few steps and pickups, the environment's panorama and operational details were recorded together, with the timing of both perspectives synchronized to sub-millisecond accuracy.

(Image source: On-site photography by Leitech)

Readers well-versed in embodied AI would understand that robots require vast amounts of training data to become intelligent. However, collecting such data typically necessitates expensive robotic hardware to perform repetitive operations, leading to a deadlock. If the robot isn't smart enough, it can't collect good data; without good data, the robot can't become smart. As a result, the global embodied AI industry's total volume of high-quality real-world interaction data is less than one twenty-thousandth of that of leading large language models.

For instance, GPT-5 was trained on 100 trillion tokens, while the entire embodied AI industry can only muster 500,000 hours of data.

Mifeng's solution is to leverage 'humans' as the primary agents for data collection. The MEgo Gripper achieves trajectory reconstruction accuracy of 1 millimeter, surpassing the previous industry standard of centimeter-level accuracy. Combined with sub-millisecond global wireless time synchronization, multimodal data such as vision, touch, and posture are precisely aligned in the same time dimension. Furthermore, its native isomorphic design with the Zhiyuan Elf G2 Air ensures that collected data can be seamlessly deployed to real robots, avoiding the issue of 'collecting data that can't be used.' At WAIC, Shen Yujun, Chief Scientist of Ant Lingbo, also noted that non-robotic data collection scales much faster than teleoperation robots, stating that 'in just half a year, several companies have already reached the 100,000-hour level in data collection.'

(Image source: On-site photography by Leitech)

The booth's large screen also showcased the MEgo Engine data governance platform in real-time, supporting one-click upload of raw data and automatically completing time alignment, 6D trajectory reconstruction, quality assessment, and intelligent labeling. This improves efficiency over traditional manual labeling by more than tenfold, with labeling accuracy reaching 99.7%.

At the main forum of the World Artificial Intelligence Conference, Yao Maoqing, Partner and Senior Vice President of Zhiyuan, President of the Embodied Business Division, and Chairman & CEO of Mifeng Technology, stated, 'The essence of a world model is an AI system capable of understanding the operating rules of the physical world, with the most crucial function being the ability to predict the next state of the world.' Therefore, to build a good world model requires mastering three key capabilities: first, multimodal fusion understanding to comprehend environmental changes; second, grasping physical laws, including dynamics and spatiotemporal understanding; and third, causal reasoning, with the ability to perform long-term reasoning without collapse.

Additionally, Yao believes that the essence of technology is a versatile, multimodal large model natively oriented toward the physical world. Nearly all data must come from the real world, meaning robots must learn the rules of the real world using real-world data.

This statement perfectly aligns with Mifeng Technology's current development trajectory: 10 million hours in 2026, scaling to hundreds of millions of hours by 2027, and striving for 10 billion hours by 2030. The core team will remain at one to two hundred people for standard prototyping, with scaled production achieved through franchised factory partnerships. In terms of pricing, while domestic real robot data currently trades at 500-1000 RMB per hour, Mifeng expects non-robotic data to cost just one-third of that.

(Image source: On-site photography by Leitech)

At this WAIC, everyone is showcasing 'what robots can do,' but only Mifeng is demonstrating 'how robots learn to perform tasks.' What embodied AI lacks is not algorithms or hardware, but massive amounts of standardized, high-quality training data. Whoever pioneers this path first will become the most indispensable link in the industrial chain (supply chain).

If embodied AI is a gold rush, then Mifeng is selling the shovels most needed in this era.

WAIC 2026, themed 'Intelligent Partners · Creating the Future Together,' is in full swing!

The AI narrative has shifted from model parameter stacking to Agent productivity deployment; heterogeneous collaboration and photonic computing continue to push computational limits; embodied AI accelerates applications, with robots entering homes and factories, making physical AI a reality.

The Leitech WAIC exploration team has arrived in Shanghai to witness the annual peak moment of AI industrialization—stay tuned for more updates!