RealMan Builds 150-Robot Data Training Ground in Changzhou to Tackle Embodied AI Data Challenges
Ji Haifeng, solution director at RealMan Intelligent Technology, provided a detailed breakdown of the current state of embodied AI data training at the 2026 Embodied Perception Fusion and Multimodal Large Model Innovation Symposium. He highlighted four recurring pain points in the industry that are frequently mentioned.
The first is the scarcity of high-quality real-machine data. The second is the high cost of collecting such data, with a single robot costing up to hundreds of thousands of yuan. The third is the gap between simulation environments and the real physical world. The fourth is the low level of industry standardization, with national standards still missing.
RealMan has taken actions in these areas, including building data collection facilities, participating in standard-setting, and developing data collection methodologies.
In May 2026, RealMan completed a data training ground in Changzhou equipped with 150 robots, open to the public. The facility is divided into two main functional areas: a basic action training zone for grasping, picking, and placing, similar to how a child learns basic motions; and a scenario application zone that sets up diverse business scenarios like home, industrial, and retail for robots to perform practical tasks.
Ji noted that real-machine data collection is not a simple one-time execution. It requires generalization across multiple dimensions, including color, objects, and actions. This process is akin to a child learning to walk and grasp objects, requiring repeated practice. The data collection follows the MCAP standard, which sets specifications for six key characteristics: time synchronization, data integrity, data accuracy, data consistency, data usability, and the difficulty of collection actions.
RealMan has participated in the drafting of multiple data-related standards and is also promoting the formulation of corresponding national standards. This aligns with the industry's need for standardization.
Data volume is a significant but often overlooked issue in training ground operations. Ji revealed that a single robot can generate 2GB of raw data per minute, and even after compression, the data volume still reaches 0.8GB. This massive volume dictates that the training ground's underlying architecture must treat storage and networking as infrastructure layers. Storage modes are divided into local and cloud storage. Cloud storage requires strict network conditions, including dedicated lines, to upload data to the cloud. Its advantage is convenient delivery, allowing direct migration of data to customers.
The MCAP data format contains various types of data. Image data is divided into compressed and raw images. In addition, there are camera calibration data, joint posture data, velocity data, pose data, end-effector-related data, and six-dimensional force data. Ji particularly emphasized time synchronization. The robot system has multiple data streams from cameras, robotic arms, and posture actions. It is essential to ensure that all data types start recording at the same time point and within the same time window. A model R&D company that purchases RealMan's data has required time deviation to be controlled within 5 milliseconds.
Remote operation is a compromise solution. RealMan found that the model of data collection relying on robots within closed venues has shown certain drawbacks, similar to the limitations of keeping a child indoors. The company advocates deploying robots in real scenarios to collect operational data. The logic of remote operation is that operators remotely teleoperate robots to complete tasks at factory sites. While performing the task, data collection is also completed. The business operation is the primary goal, and data collection is an incidental output that does not interfere with actual business operations. Ji divides the development cycle of this model into three stages. The first stage is remote operation. The second stage achieves a data closed loop, gradually reducing and even eliminating human intervention. The final stage realizes autonomous evolution of robots.
RealMan's robotic arms have achieved 50,000 hours of failure-free operation, and the company is aiming for 80,000 hours this year. Starting with ultra-light humanoid robotic arms, the company has self-developed complete robots, mainly wheeled robots, and has no plans for legged robots. Deployment cases include providing robot bodies and supporting data to a domestic research institute, providing data support to Ant Lingbo, and exporting data to a Japanese research institute. Ji has a judgment on cost: at present, a single robot is priced at hundreds of thousands of yuan. To achieve home popularity and large-scale factory applications, the robot cost needs to drop to the tens of thousands of yuan level. Whether this cost target can be achieved is directly related to whether the data training ground can continuously produce high-quality data and improve collection efficiency.