As robots evolve from the "mass production" stage to the "mass sales" stage, Qizhi Openmind aims to fill the missing middle layer of industrial embodied intelligence.
Robots seem to be one step closer to real production environments.
As a platform that centrally demonstrates robot capabilities, at this year's World Robot Conference, a growing number of enterprises have directly built specific scenarios such as factories, allowing robots to complete tasks including sorting, assembly, and inspection on site. This conference also launched a Procurement Day for the first time, where central state-owned enterprises bring real scenarios and procurement demands to look for suppliers, sending a clear signal that the industry's focus has shifted to whether robots can continuously, stably and cost-effectively create value.
However, booth demonstrations are not completely equivalent to actual implementation on production lines. Completing a task once in a relatively controllable environment, and continuously coping with workpiece deviations, environmental changes and equipment anomalies at real industrial sites, there are still multiple thresholds including data, process, reliability and engineering delivery in between.
To bridge this gap, Qizhi Openmind also brought its own solution at the conference. Spun off from EFORT, a listed intelligent robot company, this enterprise launched two new products: the HALO Skill Suit is responsible for converting people's real operation experience into trainable data, while the composite robot verifies how the same set of capabilities can complete the closed loop from perception, planning to execution at industrial sites.
Different from most enterprises that focus on the operation capabilities of robots in factory scenarios, Qizhi Openmind tries to solve the problem of how embodied intelligence can further complete training, deployment and reuse.
From Rule-Driven to Data-Driven
In fact, industrial embodied intelligence has been accompanied by doubts from the very beginning. After all, industrial robots are no longer new things on manufacturing production lines for automobiles, 3C, lithium batteries and other sectors. After decades of development, robotic arms, paired with transmission equipment, tooling fixtures and machine vision, have been able to continuously complete a series of tasks such as welding, handling, spraying and assembly.
However, what industrial embodied intelligence is really trying to enter is the "remaining part" that has not been cost-effectively covered by the automation system.
A relevant person in charge of Qizhi Openmind told 36Kr that traditional industrial robots are rigid and rule-driven: as long as the workpiece, position and environment remain unchanged, the robot can perform tasks repeatedly. But this rule-driven model cannot cope with flexible operation scenarios. Shifting from rule-driven to data-driven to improve the generalization ability of industrial robots has become the core value of embodied intelligence entering industrial scenarios.
However, the person in charge also admitted that there may be a long-standing "impossible triangle" for industrial embodied intelligence, that is, it cannot meet the requirements of fast response speed, high execution accuracy and strong cross-scenario generalization ability at the same time. This is because the more scenarios the model adapts to, the more complex the reasoning process usually is, making it harder to balance response speed and execution stability.
Therefore, in the short term, a more realistic path for the implementation of industrial embodied intelligence is to first constrain the problem within a relatively closed scope, and train vertical capabilities for a certain type of process. By moderately narrowing the generalization boundary, we can obtain faster response speed and more reliable execution results, first meet the industrial scenario's demand for robot operation stability, and then gradually expand the robot's capability scope.
Further breakdown shows that the construction of vertical capabilities for industrial embodied intelligence still faces two practical thresholds.
The first one is data. As the limitations of laboratory teleoperation and simulation data become prominent, the value of data from real production processes has become increasingly prominent. However, such data is mostly held by end customers or system integrators with industry experience, and often involves production processes, equipment parameters and trade secrets, making it difficult to be opened directly to the public, which ultimately aggravates the shortage of effective data required for training.
The second one is process knowledge. The aforementioned person in charge of Qizhi Openmind mentioned that there was a project team that repeatedly supplemented data and retrained the model after the robot had abnormal movements, but the problem was still not solved. It was not until reminded by people familiar with the process that they found the root cause was simply that the ground wire was not connected. In other words, robot manufacturers may be familiar with the ontology, motion control and models, but may not understand every specific process. Therefore, the implementation of industrial embodied intelligence also requires the joint participation of process and on-site engineering experience.
It can be seen that to truly realize the implementation of industrial embodied intelligence, what is lacking is not just a more powerful model, but also a complete toolchain that can collect data from real operations, convert process experience into robot skills, and further deploy them to real machines for execution. This is exactly the intermediate layer that Qizhi Openmind is trying to fill.
Building a Platform Layer Between Ontology and Scenarios
Following the underlying logic mentioned above, the two new products released by Qizhi Openmind this time fall at the two ends of the formation of industrial embodied intelligence capabilities respectively: the HALO Skill Suit solves the problem of how real data is collected, while the composite robot verifies how these data and model capabilities can be deployed to real machines to complete tasks in specific industrial scenarios.
According to Qizhi Openmind, the HALO Skill Suit can synchronously collect multi-modal data including 8K panoramic images, first-person perspective RGBD information, hand haptics, 27-node inertial motion capture and surface electromyography. Compared with simply collecting videos, these data of different modalities can more completely restore the movement process of people in real operations, as well as the force and physical state used when completing movements, providing training materials closer to the physical world for robots to learn fine manipulation.
In terms of engineering improvement, the new generation HALO Skill Suit features a "wear-integrated" design, which deeply integrates the whole-body sensing units into the suit body without exposed wiring. It not only supports quick putting on and taking off, fits firmly and is easy to clean and maintain, but also truly realizes "collect data as soon as you put it on". While greatly improving the wearing experience and the consistency of multiple batches of data, it also creates conditions for large-scale collection of real industrial data.
Since the collection process is not bound to a specific robot, these data can also be used for model training and skill development of robots of different brands and configurations.
However, data collection is only the first step. Human movement data needs to be cleaned, processed and trained to be converted into skills that robots can understand and execute, and then adapted to specific ontologies and industrial scenarios.
To connect the above links, Qizhi Openmind has built a complete general technical base for intelligent robots. The HALO Skill Suit is responsible for collecting real-world data, HumanGPT provides model capabilities, the Dayan data platform undertakes data management and processing, the Motodou IDE is used for skill development and debugging, and Openmind OS connects upper-layer applications with different robot ontologies, thus covering the complete process from data collection, model training, skill development to real machine deployment. On this open base similar to "Android for robots", developers can develop robot skills for specific processes, and then adapt them to robot ontologies of different brands and configurations.
The composite robot provides an execution-side verification carrier for this toolchain. It combines a robotic arm, 3D vision and a mobile chassis: the mobile chassis moves between different stations, the vision system identifies the workpiece and its position, and the robotic arm then completes grasping, sorting and assembly. Compared with robotic arms that can only repeatedly execute predetermined trajectories at fixed stations, composite robots need to continuously cope with changes in their own position, operation objects and surrounding environments, so as to adapt to more flexible production tasks.
At present, this composite robot has undergone long-term verification on the real production line of a leading automaker, for integrated operation of high-density sorting and precision assembly.
"In terms of product form, the composite robot is not an underlying innovation, what we show more is its openness in the process from development to deployment," the person in charge of Qizhi Openmind told 36Kr.
This sentence also clarifies Qizhi Openmind's positioning of itself in the industrial embodied intelligence industry chain: As a tool platform, it encapsulates underlying capabilities such as data collection, robot control and task orchestration into the development environment, so that partners with process knowledge can develop vertical applications according to actual needs. "Our idea is that everyone can develop robots, and everyone can build their own robots for vertical scenarios."
Division of Labor May Be More Important Than Full-Stack Self-Development
Qizhi Openmind's choice to position itself as a tool platform is inseparable from EFORT's long-term accumulation in the field of industrial robots.
The predecessor of Qizhi Openmind was the intelligent project team established inside EFORT in 2017, and it was not formally established as an independent company until 2024. Therefore, this is not a startup team that enters the industrial embodied intelligence sector from scratch. It has financial support from EFORT as a mature company, intelligent robot projects from EFORT, and technical accumulation precipitated by scientific research projects such as national key R&D plans and major special projects.
In addition to financial and technical support, EFORT also provides industrial sites for capability verification. At present, EFORT produces about 20,000 robots every year, with cumulative shipments exceeding 100,000 units, and it is also connected with end customers and more than 400 system integrators behind it. How these robots enter factories, how to complete secondary development around specific processes, and how to generate actual value in the end, undoubtedly constitute the starting point for Qizhi Openmind to observe the industrial automation industry chain.
This historical evolution has also shaped Qizhi Openmind's judgment on the industrial division of labor of industrial embodied intelligence. In the interview, the relevant person in charge of Qizhi Openmind frankly told 36Kr, "The full-stack self-development emphasized in the current embodied intelligence sector is essentially a manifestation of unclear industry division of labor." Algorithm teams may not understand on-site engineering problems, robot manufacturers can hardly master the processes of every industry, and it is also very costly for system integrators who understand processes to make up for capabilities such as models, motion control and data processing from scratch.
In the long-term development process, traditional industrial robots have formed a professional division of labor among core component suppliers, complete machine manufacturers, system integrators and end customers. Industrial embodied intelligence also needs to form a new collaborative relationship: robot enterprises provide ontology and execution capabilities, platform enterprises provide data, models, operating systems and development tools, system integrators use process knowledge to develop vertical applications, and end customers provide real demands, standard operation procedures and continuous operation data.
According to the plan, Qizhi Openmind will start with the switching and adaptation of the underlying system of EFORT robots, and promote more robot manufacturers to join the platform ecosystem. The ultimate goal is to realize independent control in the industrial embodied intelligence era through a self-developed industrial embodied intelligence development base that can adapt to robots of different brands and configurations, so that developers do not have to be bound by a single robot brand or a single software and hardware architecture.
However, it also needs to be noted that from completing verification within the EFORT system to truly becoming an industry-wide platform, Qizhi Openmind still needs to prove that its base can win the trust of other robot manufacturers, system integrators and end customers, and form a sufficiently rich vertical application ecosystem.
This year's World Robot Conference coincided with the listing of Unitree, the "first stock of humanoid robots", on the Sci-Tech Innovation Board. Its market performance in the early stage of listing undoubtedly reflects that the capital market still has high expectations for the embodied intelligence and robot industry. However, compared with the optimistic long-term imagination of the capital market, the actual entry of robots into production environments still needs to be promoted step by step from specific links such as data, models, ontologies and engineering delivery.
As the person in charge of Qizhi Openmind said: The mass production of robot ontologies is only the first step of industrial development. Whether more developers who understand scenarios and processes can participate and form sufficiently rich vertical applications that can create actual value determines whether robots can further move from mass production to mass sales.