From "Moving Machines" to "Autonomous Agents": Embodied Robots and Intelligent Machine Tools—A Practice of Mutual Industrial Enlightenment

From "Moving Machines" to "Autonomous Agents": Embodied Robots and Intelligent Machine Tools—A Practice of Mutual Industrial Enlightenment

Yang Yang

Member, National Expert Committee on Intelligent Manufacturing

Chairman, Hunan Zhongnan Intelligent Equipment Co., Ltd.

Intelligent machine tools and embodied robots share the same development paradigm: the shift from "pre-programmed execution" to the closed loop of "perception–learning–decision–execution," and from standalone products to systems, ecosystems, and data assets—for robotics and industrial embodied intelligence, it is essentially the same paradigm. The relationship can be viewed in both directions: what insights can the intelligent machine tool industry offer to embodied robots, and which of the experiences already proven in the embodied robot industry can serve as a reference for the development of the machine tool industry.

I. Insights from the Development of Intelligent Machine Tools for the Embodied Robot Industry

1. Do not merely build "moving machines"; build "autonomous agents." The core leap in the development of intelligent machine tools is the shift from "giving a fish" to "teaching how to fish"—from pre-set G-code (autonomous perception, autonomous learning, autonomous decision-making, autonomous execution) and factory-locked configurations that grow ever smarter with use. Mapped onto the robot industry: conventional industrial robots depend on teach pendants, scripts, and fixed workstations, whereas embodied robots are trending toward perceiving workpiece offsets, sensing forces, hearing acoustic signatures, and reading visual data under real working conditions, adjusting their own grasping strategies, assembly sequences, and motion trajectories. The acceptance criteria for embodied intelligence products should evolve from "whether fixed motions can be completed" to "whether tasks can be sustained under disturbance, incoming-material variation, tool wear, and changing illumination."
2. The "device–edge–cloud" architecture matters more than any single robot. The key to the development of intelligent machine tools lies in: the device layer (CNC units, servo drives, sensors); the edge layer (real-time inference within the machine or the workshop); and the cloud layer (model training, knowledge consolidation, and cross-machine reuse of experience). The pain point of the embodied robot industry at this stage is that "the individual unit impresses, but the fleet underperforms": one humanoid robot can perform kung fu, yet fifty robots on a stage will run into task contention, map conflicts, and unshared skills. The lesson for embodied robots is to build early: low-latency control on the device; process and force-control inference at the edge; cross-embodiment skill libraries and simulation-based training in the cloud. Otherwise, "islands of information" will simply become "islands of robots."
3. Data is not about "having some," but about "full lifecycle + command-domain/action-domain alignment." The development of the intelligent machine tool industry emphasizes: the machine tool digital thread—design, manufacturing, assembly, commissioning, and operation & maintenance; and the machining digital thread—programming, simulation, machining, and measurement—with data aligned to the "command domain" to form a closed loop of "data–samples–models." The robotics and embodied-intelligence counterparts are: the task domain—natural language, work orders, CAD, and process cards; the action domain—joints, torques, contact forces, and grasp outcomes; the environment domain—point clouds, images, acoustics, temperature, and workpiece deformation; and the feedback domain—success or failure, rework, downtime, and yield. The lesson for embodied robots is that we must not merely collect "video + motion" data in the future, but data that carries physical causality and task outcomes. Otherwise, the model learns to perform, not to work.
4. Breaking through vertical scenarios first is more realistic than the "general-purpose robot." The report stresses that intelligent machine tools should first land in: aerospace integral impellers (blisks); new-energy-vehicle motor housings and battery trays; and five-axis machining centers, turn-mill machines, and heavy-duty machine tools. The same holds for the robot industry: do not fantasize about one robot running an entire plant—start with machine tending, sorting, screwdriving, grinding and polishing, quality inspection, patrol inspection, and compliant assembly. The lesson for embodied intelligent robots: building them around "process packages / skill apps" (an assembly package for automotive motor housings, a polishing package for 3C enclosures, an inspection package for aero-engine blades) generates revenue far more readily than "putting a general-purpose VLA model directly on the job."
5. We must accelerate the building of an open ecosystem for industrial embodied robots: hardware interfaces, data interfaces, and development platforms should all be open. The machine tool industry is moving toward a more open ecosystem: open hardware interfaces; open data interfaces; intelligent functions delivered as apps; and co-creation among machine tool builders, user plants, and third parties. The biggest pain point for embodied robots today is: closed-source robot OEMs; algorithm vendors unable to access real-machine data; integrators rebuilding everything for every project; and dexterous hands, six-axis force/torque sensors, controllers, SLAM, and PLCs all requiring customized adaptation. The lesson for industrial embodied robots: the industry needs a common foundation akin to an "NC-Link / Android / industrial-grade ROS 2"—unified force/vision/joint-state interfaces; a unified task description language; and a skill marketplace through which grasping, welding, and assembly skills can be distributed as industrial apps.
6. Embodied robots need to follow the path of intelligent machine tools—from "selling equipment" to "selling capability and service." The value of the intelligent machine tool industry compounds with data accumulation: from selling products, to selling machining capability, to selling service assurance. The embodied robot industry will take the same path; otherwise, margins on robot hardware will grow ever thinner. What truly creates value is: falling unit labor-hour costs; shorter changeover times; higher yields; continuously evolving skills; and SLA-based services such as "assembly failure rate per 10,000 units." Embodied intelligence companies may, in the future, be not "robot manufacturers" but "industrial operation-skill operators."

II. Viewing the Development of the Intelligent Machine Tool Industry Through the Evolution of the Embodied Robot Industry

1. Large-model and multimodal capabilities have matured; machine tools must "get a brain" faster. The robotics field is already doing: natural-language task assignment; vision–language–action (VLA) models; world-model simulation; and teleoperated data collection plus simulation augmentation plus real-machine fine-tuning. The machine tool industry already has AI assistants, fault diagnosis, and conversational programming, but on the whole remains confined to the "machining closed loop." The next step for the intelligent machine tool is not merely thermal-error compensation, but: "give me a part drawing and a delivery date, and it generates its own process plan, selects its own tools, schedules itself, runs its own trial cuts, and conducts its own reviews." In other words, "autopilot for the mother machine of industry."
2. The "simulation-to-reality closed loop" of embodied intelligence offers a reference for the machine tool industry. The robotics field emphasizes: Sim2Real; high-fidelity physical simulation; real-machine feedback flowing back; and the data flywheel. Machine tools are, in fact, better suited to this: cutting force, thermal deformation, vibration, and tool wear all have physical models, making them far easier to model than a humanoid robot "falling over or grasping at air." The intelligent machine tool industry may consider building virtual machine tools + virtual workpieces + virtual cutting tools; running a hundred thousand trial cuts in simulation first; then feeding real machining data back—thereby forming a "process world model."
3. The robotics industry has proven that the hardware body must be—self-reliant and under our own control. The bottlenecks exposed in the embodied robot industry: dependence on servo systems, gear reducers, six-axis force/torque sensors, dexterous hands, and joint modules; and constrained access to high-end sensors and chips. In the intelligent machine tool industry, high-end CNC systems, spindles, lead screws, guide rails, and measurement feedback components remain the weak links. It follows that embodied robots and machine tools face the same challenge: full-stack self-reliance in the "brain + cerebellum + joints + sensors + actuators" of intelligent equipment. The two industries can share: servo drives; encoders; edge AI chips; real-time operating systems; and industrial communication buses.
4. Opening up scenarios matters more than "performance." The National Development and Reform Commission and the World Robot Conference in Beijing E-Town have alike stressed: opening factory workstations; building real-scenario training grounds; running tasks on real machines; and evaluating by "task success rate, takeover rate, MTBF, and payback period." The machine tool industry has long been caught in the loop of "prototype–appraisal–award–failure to scale." Intelligent machine tools likewise need "industrial real-scenario proving grounds": running continuously for three months in aero-engine plants, battery-tray plants, die and mold shops, and reducer plants—encouraging yields, cycle times, tool-change counts, and downtime durations to do the talking.
5. Standards should be set early—but must not freeze innovation. The robotics field has begun defining: a humanoid robot standard system; dexterous-hand interfaces; data formats; and safety and ethics. Machine tools likewise have NC-Link, umati, and MTConnect, but adoption remains limited. For the intelligent machine tool industry, standards should first cover "interfaces / safety / evaluation / data dictionaries"; there should be no rush to lock down "intelligence levels, model architectures, or application forms"—leave ample room for the evolution of device–edge–cloud, app ecosystems, and domain-specific large models.

III. Insights from the Development of Intelligent Machine Tools for Embodied Robots

Because they share the same paradigm—"perception–cognition–decision–execution–learning–collaboration–evolution"—we can no longer treat robots as automata; they must be built as "industrial physical agents": device–edge–cloud collaboration, full-lifecycle data, vertical process closed loops, open app ecosystems, ever smarter with use, and selling capability rather than iron. And the lesson the embodied robot industry offers to intelligent machine tools is: do not remain merely "machines that cut"—transform, as quickly as possible, into "industrial embodied intelligent agents that understand the language of process, learn by trial and error in simulation, evolve on the production line, and work in coordination with robots and AGVs."

The factory of the future is not an assemblage of "CNC machine tools + robots"; rather, the intelligent machine tool becomes an embodied machining agent; the industrial robot evolves into an embodied manipulation agent; AGVs and automated storage and retrieval systems evolve into embodied logistics agents; and MES together with large models converges toward a collective-intelligence brain for the plant.