In embodied intelligence, few points draw as much agreement as the importance of data.

Large language models such as OpenAI’s GPT were trained on vast text corpora, and the so-called scaling law, the idea that model performance improves predictably with more data and compute, has become a working assumption across much of artificial intelligence. In the physical world, however, there is no equivalent reservoir. Whether human or robot, the volume of real-world interaction data remains far below what would be required to recreate a GPT-like inflection point for embodied systems.

That gap raises practical questions. How is the data collected? At what scale? And how is quality maintained? For companies building embodied intelligence, the answers increasingly shape their prospects.

Lumos, a robotics startup founded in September 2024, has chosen to focus on the first link in that chain: data collection.

The company has launched what it describes as a backpack-mounted universal manipulation interface device called FastUMI Pro. It plans to deploy 10,000 units in 2026 across six real-world environments, industrial sites, homes, hotels, restaurants, shopping malls, and offices, to conduct systematic data collection.

To understand the strategy, it helps to clarify what UMI is. Proposed jointly by researchers at Stanford University, Columbia University, and the Toyota Research Institute, UMI is a low-cost data collection and learning framework. Unlike teleoperation-based methods, which typically bind data to a specific robot’s hardware, UMI decouples the acquisition system from the robot itself. In principle, that allows the resulting data to generalize across different robot morphologies rather than remaining tied to a single machine.

At a media briefing earlier this year, Lumos founder and CEO Yu Chao compared UMI’s efficiency and cost with teleoperation.

“For a task like folding clothes, teleoperation data collection takes about 50 seconds and costs RMB 3–5 (USD 0.42–0.70),” Yu said. “Using FastUMI Pro, it takes just ten seconds, and costs less than RMB 0.6 (USD 0.08). That significantly improves efficiency while lowering costs.”

Yu previously led embodied robotics at Dreame, where he oversaw the development and mass production of Xiaomi’s CyberDog, delivering more than 1,000 units, according to the company. Co-CTO Ding Yan, who worked on UMI in its early research phase, was among the first in China to push the framework toward industrial applications.

In 2025, Lumos built a dedicated data collection center and said it achieved an annualized production capacity of 100,000 hours of data. Yu estimates that by 2026, leading embodied AI models will require at least one million hours of training data.

Lumos’ stated goal for 2026 is to achieve an annual UMI data production capacity of one million hours. Reaching that scale would require moving beyond centralized facilities toward distributed collection in everyday settings.

“Robot training data should not be this expensive or scarce. Humans generate data constantly while working in the physical world. It’s everywhere. It just hasn’t been properly collected.”

The FastUMI Pro is designed around that premise. Worn as a backpack, it functions as a portable, standardized workstation that converts real-world operations into structured training data. Historically, embodied data collection has relied on laboratories or tightly controlled environments. That approach often produces repetitive datasets, with robots performing a narrow range of actions in fixed contexts, limiting a model’s ability to generalize.

By miniaturizing its toolkit, Lumos is attempting to lower the barrier to capturing more diverse operational data.

The company plans to target six broad scenarios, industrial, residential, hospitality, food and beverage, retail, and office environments, subdivided into 30 task categories. It describes the objective as building a structured, multidimensional operational dataset.

At the heart of Lumos’ system is what it calls an integrated loop connecting collection, training, and deployment. The new initiative relies on that closed cycle.

Using FastUMI Pro, Lumos said its dual-arm robot Mos completed a factory quality inspection workflow, including data collection, policy training, and model inference, within five hours. After on-site deployment in Hefei, the company said the system completed the same sequence in seven hours in a real-world setting.

In parallel, Lumos has introduced what it calls a “data supermarket.” It has standardized portions of its datasets into products that customers can purchase through its website.

For now, Lumos’ strategy centers on data infrastructure rather than model architecture.

“I come from a model training background,” Ding said in interviews with several media outlets earlier this year. “I used to focus on training models. But we ran into a major issue. To train a really strong model, you need a solid data pipeline including data production, evaluation, and filtering. Building that pipeline takes time.”

After evaluating its priorities, the company decided to focus first on data.

“In the end, competition in model architecture only goes so far,” Ding said. “What truly differentiates models is the quality of their data.”

Whether that view proves durable as embodied AI evolves remains uncertain. What is clear is that the upper bound of embodied intelligence is constrained by the scale and diversity of real-world operational data. If standardized datasets can be sourced as readily as hardware components, the threshold for training industry-grade models may decline.

Lumos is betting that shifting from centralized labs to thousands of backpack-mounted devices will expand that supply. If operational data becomes more abundant and standardized, embodied systems may begin to generalize beyond controlled demonstrations and into routine, repeatable work.

KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Qiu Xiaofen for 36Kr.