36Kr Exclusive | A team with Tencent and Baidu backgrounds dedicated to building embodied data infrastructure has secured a new round of financing of several hundred million yuan from Shunwei Capital and other investors.
Author | QIAO Yujie
Editor | YUAN Silai
36Kr learned that IO-AI Tech, an enterprise dedicated to robotics and embodied intelligence data infrastructure, has announced the completion of a financing of several hundred million yuan. This round of financing is jointly invested by Shunwei Capital, Songhe Capital and Shenzhen Venture Capital, with strategic investment from leading robot ontology enterprises.
Founded in 2023, IO-AI Tech is a technology company focusing on robotics and embodied intelligence data infrastructure, committed to building a bridge connecting data, robots, models and real scenarios to support the continuous improvement of robot capabilities and large-scale application.
The founding team comes from top technology and robotics enterprises such as Tencent, XPeng Motors, ByteDance, Amazon and Baidu, with profound technical accumulation and industry experience in the fields of robotics and AI. Chen Xiangyu, Founder and CEO, holds a doctorate from the University of Tokyo, has won the ICRA Best Paper Award twice, and once served as the project leader of Tencent's quadruped robot project. Gao Biao, Co-founder and CTO, holds a doctorate from Peking University, and once worked as a senior algorithm engineer of Baidu Apollo Go.
As embodied intelligence gradually moves from the technical verification stage to industrial implementation, the importance of robot training data is increasing. Unlike large models that mainly rely on data such as Internet text and images, robots need to understand the spatial relationships, action processes and interactions between humans and objects in the real physical environment, and continuously improve their capabilities by obtaining real-world feedback, while there is still no mature and efficient way to obtain these data at present.
Focusing on this demand, IO-AI Tech has built a data infrastructure system covering data collection, data management and model training, and has now formed three product lines: the TeleXperience teleoperation system, the SenseXperience real-world human data collection system and the EmbodiFlow data management platform.
Image source / The enterprise
Image source / The enterprise
Image source / The enterprise
Among them, TeleXperience is mainly oriented to teleoperation and data collection on the robot side.
The core capability of this system is to stably and low-latency map human movements, perspectives and operation intentions to robots of different configurations. At the same time, the data generated in the whole process can be recorded and entered into the subsequent annotation, management and training process, realizing the efficient transfer of human experience to robot motion capabilities, and providing a unified entry for robot training, remote operation and real task data collection. The value of TeleXperience is not only to control robots, but also to help customers quickly build a closed loop of "task execution - data collection - model iteration".
In 2025, IO-AI Tech participated in the ICRA Bimanual Robot Capability Boundary Challenge with TeleXperience. As the only cross-border teleoperation participating team, it stood out from 88 teams around the world and won the Best Application Award. Up to now, TeleXperience has been adapted to more than 80 types of robots, covering various robot forms such as humanoid robots, robotic arms and mobile robots.
In addition to robot operation data, IO-AI Tech is also deploying real-world human behavior data collection.
The SenseXperience system launched by the company mainly collects human operation data in real scenarios through the first-person perspective unit, wrist camera and UMI gripper, providing data sources for embodied intelligence model training.
According to the introduction, the early version of SenseXperience adopted a relatively complete multi-modal collection solution, including sensing devices such as whole-body motion capture, vision and tactile sense. Based on this system, IO-AI Tech has previously accumulated millions of levels of human operation data and participated in the construction of relevant open data sets.
With the change of market demand, the company has made lightweight adjustments to its products. The currently launched lightweight version of SenseXperience, the Baseline Solution, reduces complex wearable devices, adopts a modular collection method, and realizes lower-cost and scalable data production through head-mounted devices, wrist collection units and gripper collection solutions.
In the data processing link, IO-AI Tech launched the EmbodiFlow data management platform, covering processes such as collection, annotation, review, quality inspection, visualization and data format conversion. The platform supports output of data formats commonly used in the robotics field such as LeRobot, MCAP and HDF5, lowering the processing threshold from raw data collection to model training.
At present, IO-AI Tech mainly focuses on the large-scale application requirements of robotics and embodied intelligence, providing productized capabilities and solution support for robot enterprises, model companies and industrial customers. On the one hand, the company provides basic capabilities such as robot teleoperation, real-world data collection and data management to robot ontology enterprises and model companies through products such as TeleXperience, SenseXperience and EmbodiFlow. On the other hand, it is also exploring to extend data capabilities to actual robot deployment scenarios.
Chen Xiangyu, Founder of IO-AI Tech, said, "In the past three years, we have focused on the data closed loop, and got through the data collection, cleaning, annotation and verification links. Next, we hope to further realize the landing closed loop, so that robots can truly enter production and life scenarios." At present, IO-AI Tech is exploring application directions such as industry, commerce, agriculture and households, and plans to further expand overseas markets.
According to the introduction, the funds raised in this round will be mainly used in three aspects: continuously increasing investment in core product R&D, building a diversified R&D team, and accelerating global market expansion to improve the large-scale application capability of robots.
The following is an excerpt of the exchange between 36Kr and founder Chen Xiangyu:
36Kr: The lightweight version of SenseXperience has entered the stage of stable mass production and delivery. What are the core considerations in the design of this system?
Chen Xiangyu: Actually, we have been working on human data collection since 2023. At that time, we didn't plan to sell equipment, but because we needed data ourselves, we started collection first. The early version of SenseXperience was actually a relatively complete solution, including multi-modal data such as whole-body motion capture, vision, and tactile sense. We also collected millions of levels of human operation data ourselves and participated in the construction of some data sets.
Later, we observed some changes in the market. In the past, people paid more attention to whether there was data, hoping to quickly expand the data scale and verify the model capability. But as robots begin to move to real scenarios, people will pay more and more attention to data quality and whether the data comes from real tasks.
So the new version we launched now has made some subtractions. In the past, the equipment was relatively heavy, but now we hope to make it more lightweight, so that the collection personnel do not need to wear complex equipment, and it will not affect their normal work. At the same time, we also hope that the equipment is more stable and can support large-scale deployment.
In addition, the industry has not yet fully converged on the recipe for robot training data at present, and different robots, different models and different scenarios require different data. Therefore, our idea is not to define a fixed solution, but to modularize the collection capabilities and combine them according to customer needs.
36Kr: At present, real machine teleoperation data is still regarded as the most valuable data source. What changes are taking place in the "recipe" of training data? How will the company balance the layout of different types of data?
Chen Xiangyu: I think the focus is different at different stages. In the early stage, scale is definitely needed, because without enough data, it is difficult for the model to generate capabilities. Many times, people will first feed the data to see if the model can learn something, which is a rapid verification process.
But in the later stage, the value of high-quality data will definitely become more and more obvious. In fact, the development of large models is similar. In the early stage, people may rely more on general data, but later they will find that high-quality, professional domain data is very important for the improvement of model capabilities.
Therefore, we will not judge that a certain data route is necessarily correct, but will continue to observe the model development and technical changes. For IO-AI Tech, in the long run, we still hope to produce more valuable data, especially high-quality data in real tasks. At the same time, we will also provide corresponding products and services for customers' data needs at the current stage. Finally, we hope to form a continuous closed loop from real-world data acquisition, robot capability optimization to application feedback iteration, so that robots can continuously learn and grow in more real scenarios.