The vision for robots in manufacturing is being redefined.
Over the past three decades, industrial robots have addressed the question of "whether automation is possible." Traditional industrial robots have long relied on manual programming, fixed tooling, and repetitive trajectories, making them suitable for highly structured production scenarios with clear task boundaries. Essentially, they are specialized equipment designed for "fixed scenarios, fixed processes, and fixed programs," requiring dedicated hardware adaptation and deeply customized programming for individual workstations. With high changeover costs and limited adaptable scenarios, they serve standardized mass production.
However, China's manufacturing sector is shifting from large-scale standardized production to a flexible model characterized by high-mix and low-volume production.
Statistics show that high-mix and low-volume production enterprises currently account for over 65% of China's discrete manufacturing industry, with product changeover cycles shortened by an average of about 30%. Meanwhile, the relatively long changeover and debugging cycles and high retrofitting costs of traditional robots are increasingly mismatched with this flexible demand.
It is precisely at this industrial turning point that general-purpose intelligent robots have entered the industry's field of vision. This does not refer to a specific form of hardware, but rather a class of robot systems capable of cross-industry and multi-task operations. Through a closed loop of perception, decision-making, and execution, general-purpose intelligent robots can quickly adapt to new tasks with minimal teaching or even natural language commands.
How dynamic is this sector today?
According to data from MIR Rui Industry, the market size of complete general-purpose intelligent robots in China will reach approximately CNY 8.4 billion in 2025. Among them, the intelligent collaborative robot market will be about CNY 2.5 billion, with sales exceeding 40,000 units, a year-on-year growth rate of over 30%; the humanoid robot market size will be about CNY 5.1 billion, expected to increase to CNY 12.5 billion in 2026.
Image source: "2026 White Paper on the Development of China's General-Purpose Intelligent Robot Industry"
Beneath the hype, however, the industry has not yet reached a consensus on "what exactly general-purpose intelligent robots are, how far the technology has progressed, and how close we are to true generalization." On October 13, MIR Rui Industry officially released the "2026 White Paper on the Development of China's General-Purpose Intelligent Robot Industry" (hereinafter referred to as the "White Paper"), systematically reviewing this sector from dimensions such as industry definition, industrial chain structure, market landscape, technology routes, and development trends, providing a reference for industry participants.
01. Boundaries and Lineage of General-Purpose Intelligent Robots
Compared to traditional industrial robots, general-purpose intelligent robots are robotic systems equipped with environmental perception, autonomous decision-making, and learning evolution capabilities. They can perform multiple types of tasks in unstructured environments and adapt to diverse scenarios relying on modular and reconfigurable hardware and software architectures. Distinct from traditional specialized robots designed for single tasks and fixed scenarios, their core characteristics are task universality and scenario adaptability, allowing flexible deployment in industrial, commercial, and special fields.
In terms of physical form, general-purpose intelligent robots are divided into two main categories: fixed robots and mobile robots. Fixed robots, with collaborative robots as the core category, include single-arm and dual-arm collaborative forms, featuring mature technology and reliable precision. Mobile robots possess dual capabilities of autonomous mobility and operation, further subdivided into single-arm wheeled robots, wheeled humanoid robots, and legged humanoid robots.
In such a sector, the players rushing in are more numerous and complex than in the traditional robot industry.
Midstream manufacturers are roughly divided into two camps:
On one side are the robot body manufacturers deeply engaged in the robotics field, which can be further subdivided into four categories: the veteran forces transformed from traditional industrial robot manufacturers such as Aisidun, Aifute, and Inovance; companies extending from collaborative robots to intelligent collaboration and humanoid forms, such as Jieka, Rokae Robotics, and Yuejiang, among which Jieka has built a product matrix covering single-arm, dual-arm, wheeled humanoid, and legged humanoid forms; manufacturers extending from bionic robot companies such as Yushu and Yunshenchu to collaborative and general-purpose robots; and the new humanoid forces such as AgiBot, Yinhetongyong, and Mofayuanzi, which broke into the field since 2023 with a compound background of "AI + robotics."
On the other side are the industrial forces crossing over from other fields: in the automotive camp, Tesla's Optimus has initiated the mass production process, and BYD's "Yaoshunyu" robot project, XPeng's IRON humanoid robot, and GAC's Huilun have entered the market one after another, while Tier 1 automotive component suppliers such as Tuopu Group and Junsheng are also cutting into core components; in the technology camp, Xiaomi has incorporated robots into the "smartphone + automobile + AIoT" strategic framework, and Ant Group is deploying embodied large models through Robbyant.
According to the autonomous operation capability from low to high, general-purpose intelligent robots are divided into four levels from L2 to L5:
Evolution Chart of Autonomous Operation Capability of General-Purpose Intelligent Robots (L2-L5)
Image source: "2026 White Paper on the Development of China's General-Purpose Intelligent Robot Industry"
Currently, the global industry is in a critical stage of large-scale transition from L3 to L4. L3 products possess both implementation feasibility and commercial value, serving as the main force in industrial scenarios over the next two to three years. Among general-purpose intelligent robots, collaborative robots are the most commercially mature and highest-penetrating L3 form at this stage. Relying on joint torque control to achieve human-robot integration, they can operate at the same workstation with humans without safety fences, featuring flexible deployment. With the support of AI, they can autonomously complete workpiece recognition, trajectory planning, and exception handling. In contrast, although the humanoid robot market is large in volume, its applications remain concentrated in scientific research and education, cultural and entertainment performances, and demonstration production lines, with a limited number of actual units in mass production workshops.
Image source: "2026 White Paper on the Development of China's General-Purpose Intelligent Robot Industry"
Collaborative robots have mature bodies, an existing customer base, and real-scenario data feedback, which are precisely the most needed soil for training an intelligent robot that can actually work.
02. Technology Routes of General-Purpose Intelligent Robots: Hierarchical Modularization or End-to-End VLA?
To determine whether this machine is truly intelligent enough, one must first look at how its brain thinks.
Currently, the industry is actually running on two paths.
One is the hierarchical modular architecture, represented by the engineering solutions of leading Chinese manufacturers such as Jieka. The system is divided into four levels: perception module, large model cognitive module, motion planning module, and underlying motion control module. Each module is independently developed and iterated in a decoupled manner, interacting and communicating through standardized interfaces.
The characteristics and advantages of this architecture are stronger stability in engineering implementation, easier module-by-module debugging and troubleshooting, and better adaptation to the implementation of real machines in industrial and commercial scenarios. It is also the mainstream technical choice for the embodied transformation of current collaborative robots and single-arm wheeled composite robots.
Taking the JAKA EVO industrial embodied intelligence platform as an example, it adopts a hierarchical agent architecture: the cloud large model is responsible for visual perception, task planning, skill selection, and exception decision-making; the edge model combines a rule engine to be responsible for action planning and force control execution; the underlying layer uses mature motion control algorithms to ensure stable execution with high-speed response, which not only allows continuous upgrading of the upper-level model but also avoids the model's "hallucinations" affecting the underlying safety control.
The other is the end-to-end VLA (Vision-Language-Action) model route, represented by research from institutions such as Google, NVIDIA, and Stanford University.
The core of this technology route is to directly map visual and language inputs to robot action outputs, weakening the intermediate artificially designed perception and planning sub-modules, and allowing the model to autonomously learn the mapping relationship between vision-language and actions. The end-to-end route has stronger generalization potential in complex unstructured scenarios, but requires more training data and computing power resources. At this stage, it lacks sufficient robustness and is more applied in frontier scientific research verification of humanoid robots. The open-source model OpenVLA launched by Stanford University in 2024 and NVIDIA's GrootN1 both belong to this lineage.
Regarding these two differentiated technology routes, the industry is also exploring integrated solutions. For example, GeneralVLA adopts a hierarchical design: the upper layer uses a vision-language model to complete task understanding and object manipulable position recognition; the middle layer performs knowledge-guided trajectory planning; and the underlying layer executes action control, balancing generalization ability and engineering stability.
Technology routes must ultimately be implemented on hardware and software platforms. Such platforms are mainly divided into two categories: one is the robot body hardware and software platform, covering collaborative robots and humanoid robot bodies, as well as servo joints, multimodal sensors, onboard computing units, and supporting control software stacks. Among them, the collaborative robot platform relies on mature bodies superimposed with AI capabilities and uses natural language programming to lower the deployment threshold. The Jieka Coboπ EI platform supports automatic generation and mixed execution of Python scripts and native JKS scripts, allowing users to generate operation scripts by inputting natural language task descriptions; the humanoid robot platform focuses on whole-body motion control, force perception, and safe human-robot interaction; the single-arm wheeled platform integrates the hardware and software of the mobile chassis and the operating arm.
The other category is the simulation and digital twin platform, serving as infrastructure to complete action trial-and-error and task verification without occupying real machines. Domestically, there are already Songying Technology's ORCA OS physical AI operating system and Jieka's JAKA Studio Pro simulation platform. Among them, Jieka's JAKA Studio Pro follows the Sim2Real technology route of directly deploying simulation training models on real machines. Relying on the simulation platform, it batch generates multi-scenario and multi-interference simulation datasets. The real machine only needs a small amount of real samples to complete domain transfer. Meanwhile, over 30,000 operating robots worldwide continuously collect real industrial interaction data, building an exclusive embodied dataset for industrial scenarios and forming a data flywheel.
03. Key Technologies of General-Purpose Intelligent Robots: A Full-Link Closed Loop of Training, Simulation, and Scenario Deployment
A deeper question than route selection is: where exactly do the robot's capabilities come from? The answer lies not in the body, but in an invisible backend system.
01. Training
The training data for general-purpose intelligent robots come from four categories: real teleoperation, production site operation, internet multimodal content, and simulation synthesis, with the industry generally adopting a hybrid approach. In terms of training paradigms, VA, VLA, WAM, UAM, and reinforcement learning are evolving in parallel. Among them, WAM upgrades the robot from "execute upon seeing" to "predict before acting," endowing actions with physical consequence constraints, which significantly improves the rationality and robustness of actions. This direction is conducive to enhancing the robot's generalization ability in open environments, but at this stage, it still faces problems such as slow inference speed, insufficient 3D space understanding, and difficulty in cross-body transfer.
Image source: "2026 White Paper on the Development of China's General-Purpose Intelligent Robot Industry"
02. Simulation
The role of simulation is also changing. It was used for offline programming and collision checking in the past, and has now become infrastructure for data generation, model training, and pre-deployment testing. NVIDIA once used a simulation platform to generate 780,000 synthetic trajectories in 11 hours, equivalent to about 6,500 hours of manual teaching. After mixing with real data, the model performance improved by 40%. Of course, Sim2Real remains a core difficulty. In the future, it will rely on real machine data to reverse-calibrate simulation parameters, forming a closed loop of training-verification-feedback-iteration.
03. Deployment
On the deployment end, AI natural language programming lowers the threshold for using robots from "knowing how to program" to "knowing how to describe tasks." After the user describes the task, the system automatically completes recognition, planning, and safety verification. Taking Jieka JAKA EVO as an example, it adopts hierarchical control and a 3B multimodal manipulation model. Cross-task transfer can be completed within 100 sets of teaching data and 6 hours, achieving zero-shot generalization for similar tasks, with cross-modal semantic recognition accuracy of no less than 95%. In the future, deployment platforms will evolve into an "AI Hub" that uniformly manages models, skills, data, and equipment status. Deployment will no longer be a one-time delivery, but the starting point for continuous capability evolution.
Conclusion
Objectively speaking, general-purpose intelligent robots already have a commercial foundation in command understanding, conventional grasping, simple sorting, and trajectory planning. The industry's largest public dataset gathers over 1 million real robot trajectories and covers 22 types of bodies, but compared to the massive amount of image and text data on the internet, robot interaction data remains a scarce resource. In cross-industry and significantly changing unknown environments, the accuracy of complex process operations is still less than 90%; encountering sudden changes in lighting, workpiece deviation, or unexpected occlusion, tasks are still prone to interruption.
This means that in the short term, general-purpose intelligent robots are most suitable for tasks with shorter action chains, clear goals, and relatively controllable environments. The real opportunity lies not in all-around operations in open environments, but in semi-structured scenarios such as manufacturing, logistics, inspection, patrol, and commercial services, replacing traditional automation solutions with lower changeover costs and shorter deployment cycles.
Looking forward, several evolutionary directions have become clear:
Foundation models enable the same robot to change from "one program for one task" to "one body with continuously expanding skills"; multimodal large models continue to sink to the edge, and edge-cloud collaboration balances real-time performance and intelligence;
World models and predictive intelligence become the high ground for tackling challenges in the next stage, enabling robots to learn to predict the consequences of actions;
The data flywheel from training to deployment continues to spin, and the real operational experience accumulated at industrial sites will become the hardest competitive advantage;
Safety and verifiability will become prerequisites for industrialization as robots step out of safety fences and enter human-robot shared line scenarios.
The ultimate outcome of general-purpose intelligent robots is not a single product reshaping human life, but the progress of an entire set of technological paradigms. Robots are gradually growing from equipment that repeats actions according to programs into intelligent operational units capable of understanding tasks, predicting results, and continuously learning. As China's manufacturing industry is undergoing flexible upgrading with its globally unique industrial depth, and when massive real workshop scenarios become training soil that no laboratory can replicate, what this migration can release will not only be the growth of a robot track, but the improvement of the efficiency and resilience of the entire manufacturing system.