Just Say "Clean Up": China's Robots Evolve by Absorbing Field Data

"Clean up the desk neatly."

When a simple sentence was typed into a screen resembling a Google search box, a humanoid robot as tall as a person began busily moving its gripper hands in front of a cluttered desk. Without any separate instructions, the robot recognized the items on the desk on its own and tidied them up. Even when new objects such as a watch were placed on the desk, or the work was obstructed by hand, there was no particular problem in carrying out the task. A series of execution processes such as "open the drawer" and "put the black cable in the drawer" were displayed on the monitor in real time, and once one task was completed, it moved on to the next.

On Thursday, at the headquarters of Spirit AI, located in the Zhongguancun Software Park in Haidian District, Beijing's largest information technology (IT) corporate cluster, one could meet such robots that judge and act on their own. Spirit AI is a startup that has focused on directly confronting the weakness that Chinese robots showing off flashy dances and acrobatics actually only simply repeat pre-programmed movements.

What made this possible is "Spirit v1.5," a vision-language-action (VLA) model the company unveiled in January. VLA serves as the robot's "brain," recognizing the environment with sensors and cameras, understanding and judging language and context, and then immediately proceeding to action. Its core competitiveness lies in its versatility in recognizing objects and planning tasks in unfamiliar environments without detailed commands.

Spirit AI has the strength of utilizing real-world data rather than internet photos or simulation data. This is because manipulation tasks such as fitting screws or opening locks can fail due to even minute differences in friction and clearance, making the data accumulated by actual robots in the field important. Thanks to this, Spirit AI is evaluated as able to demonstrate higher performance at manufacturing sites such as factory assembly. Manufacturing companies that require strict processes, such as Bosch and CATL, have also joined as strategic investors and are deploying Spirit AI robots in their factories.

Spirit AI has secured more than 400,000 data collection sites across China and has independently developed infrastructure to improve data collection efficiency, including remote control software and wearable equipment. On the day of the visit, an experiment with a recently developed three-finger wearable glove was underway on one side. The company explained that it can more precisely reflect the actual movements of a human hand than the existing two-finger equipment. A Spirit AI official stressed, "Our goal is to secure 1 million hours of valid data within the year."

The results are showing up in the metrics. Spirit 1.5, released as open source this January, outperformed the model of Physical Intelligence (PI), a leading U.S. VLA startup, on "RoboArena," a platform for evaluating robot brain performance. Interest from capital markets is also growing. Spirit AI surpassed a valuation of 20 billion yuan in March this year, joining the "20 billion yuan club" along with Unitree, AgiBot, and Galbot. Venture capital firms run by Xiaomi founder Lei Jun and Alibaba founder Jack Ma also participated in the investment, and the number of employees has more than doubled from around 100 to over 200 this year. The company explained that many excellent research personnel with master's and doctoral degrees from Tsinghua University have joined, centered on co-founder and chief scientist Gao Yang, an assistant professor at Tsinghua University.

The site visited on this day had more space for robots than any robot company visited so far. Of the space spanning two floors, office space accounted for only a very small portion, while data collection and refinement areas and VLA model demonstration areas occupied a much larger share. The robots were densely arranged at intervals of tens of centimeters, and a significant number of employees were also focused on accumulating and verifying robot data while wearing equipment themselves, rather than sitting at desks.

However, the model itself is not yet at a stage of generating profit. The company explained that since the current robot brain model is at an early stage, it first produces and sells robot bodies to generate revenue, and supports research and development (R&D) by installing its own model demo on them for training. A company official said, "The robot's brain is still at the baby stage, so it can only do very simple things," but added, "The key is to secure the most valid data compared to competitors."