The convergence of force control and motion control: jointly exploring breakthroughs in the evolution of the embodied intelligence “cerebellum.”

As embodied intelligence evolves from “being able to think” to “being able to act,” the underlying motion control system—the “cerebellum”—is emerging as the key determinant of whether robots can truly be deployed in industrial settings. Along the evolutionary trajectory of this embodied‑intelligence cerebellum, force control and motion control represent two highly influential technological pathways. The force‑control approach centers on tactile feedback, mimicking human fingertip sensing and compliant interaction, striving for “precise force application like a human”; the motion‑control approach emphasizes trajectory planning combined with whole-body coordination, replicating human limb coupling and balance regulation, aiming for “agile locomotion akin to human movement.” If traditional industrial robots are essentially “programmatic executors” that perform tasks by following pre‑defined waypoints, then today’s robots equipped with advanced motion‑control capabilities are already advancing toward becoming genuine “motion intelligents.”

I. Force Control and Motion Control: Complementary Technological Pathways

Force control technology focuses on force sensing and force regulation during the interaction between robots and their environment. By deploying force sensors at the joint outputs, robots cease to be rigid “hard‑collision” actuators reliant on position control; instead, they become flexible collaborative agents capable of sensing contact forces and dynamically adjusting their postures in real time. Adaptive robots integrate high‑precision force control, visual perception, and AI algorithms, endowing them with environmental adaptability and autonomous decision‑making capabilities. At their core, these systems establish a closed loop of “perception–decision–actuation,” enabling human‑like hand–eye coordination. Force control enables robots to sense and adaptively modulate applied forces during tasks such as assembly, polishing, and plug‑in/plug‑out operations, thereby preventing damage to workpieces and serving as a critical foundation for compliant manipulation.

Motion control technology focuses on trajectory planning and whole-body coordination for robots. For example, when a robot encounters a slippery surface during locomotion, its motion-control cerebellum can directly leverage force-sensor feedback to dynamically adjust the support forces of the leg joints and adapt the gait in real time, thereby preventing falls. At the same time, it synchronously transmits environmental information to the perception brain, enabling it to update path-planning algorithms and achieve parallel processing of “safe response” and “intelligent decision-making.” In industrial settings, motion-control technology ensures both the precision of multi‑joint coordination and rapid dynamic responsiveness.

The two approaches each have distinct emphases: the force‑control paradigm prioritizes compliance and force sensing, making it well suited for precision assembly, polishing, and other tasks that demand fine‑grained force regulation; the motion‑control paradigm emphasizes dynamic responsiveness and whole‑body coordination, ideal for multi‑joint collaborative movements and complex motion control. However, in real‑world industrial settings, robots often need to combine precise force application with agile locomotion, a capability that neither approach alone can fully address. Industry experts generally agree that these two technical paths are best matched to different use cases, and that modern robots must not only execute precise motions but also interpret force feedback. Advances in tactile sensing, end‑to‑end learning models, and cloud‑edge collaboration are driving robotics forward—from merely “being able to move” to “being able to manipulate.”

II. The Ongoing Evolution of “Cerebellum” Technology and Market Growth

After years of technological iteration, the field of “cerebellar” control has achieved groundbreaking advances—ranging from playing tennis and table tennis to dancing, performing martial arts, and executing complex full-body movements. Today’s humanoid robots can handle a wide array of highly sophisticated motion-control tasks. The continuous improvement in motor control capabilities is bringing robotic behavior ever closer to that of humans, while also laying the technical groundwork for their deployment in real-world settings.

On the technological front, the three major dimensions—“brain,” “cerebellum,” and body”—are undergoing co‑evolution. Large general‑purpose AI models are driving a rapid leap in the cognitive capabilities of robotic “brains,” while the localization rate of critical components such as dexterous hands, joints, and reducers has risen markedly. According to relevant industry roadmaps, by 2027 China aims to achieve major breakthroughs in embodied brain–body systems, whole‑machine control chips, and full‑body motion‑control capabilities, thereby enabling the intelligent, efficient, and large‑scale deployment of embodied intelligent robots.

At the market level, embodied intelligence is experiencing rapid growth. According to industry research firms, global revenue from integrated embodied “big‑brain” and “small‑brain” systems is projected to reach approximately RMB 1.7 billion in 2025, with estimates suggesting it will approach RMB 8.5 billion by 2032—driven by a compound annual growth rate exceeding 25%. By 2025, the humanoid robotics sector is expected to transition swiftly from the “technology validation phase” to the “large‑scale commercialization phase.” On the capital front, investment continues to ramp up, with numerous robotics companies securing substantial financing; investors widely view underlying technologies such as motion control as pivotal for bridging the critical gap between the “brain” and the physical world.

III. The Path of Integration: From “Separation” to “Collaboration”

In recent years, the industry has increasingly recognized that the deep integration of force control and motion control is the inevitable direction of “cerebellar” evolution. This integration is unfolding across multiple dimensions:

At the control‑architecture level, hybrid force–position control has emerged as a key technological direction. A multi‑sensor collaborative control system dominated by force control uses force/torque sensors as the primary sensing modality, while integrating vision and laser‑based positioning to achieve an integrated perception‑control architecture that enables precise control of contact tasks. By embedding high‑precision torque sensors in each joint axis, it is possible—without significantly increasing hardware costs—to simultaneously deliver high‑accuracy motion control and compliant force‑controlled manipulation. Torque measurement resolution can exceed 0.1 Nm, and the closed‑loop bandwidth for joint‑level force control can reach several kilohertz.

At the industrial application level, robots that employ full-body force‑control technology excel in dual‑arm cooperative impedance control, vehicle‑arm collaborative trajectory planning, and vehicle‑arm traction‑driving operations. Their torso is equipped with collision‑detection capabilities, ensuring autonomous safety protection and safe human–robot collaboration.

At the technological ecosystem level, open source and openness are accelerating the integration process. Industry research institutions have released a general-purpose robotic cerebellum‑level intelligence framework that spans from simulation training to real‑robot deployment, as well as a full‑body control framework tailored for humanoid robots. Meanwhile, VLA models designed to enable embodied cerebellar capabilities, along with large‑scale datasets that support their training, have also been progressively made open source. These open‑source contributions have effectively lowered the R&D barrier for cerebellar technologies and fostered integrated innovation in force control and motion control.

IV. Industry Challenges and Future Prospects

Despite the significant advances in “cerebellum” technology, its industrial implementation still faces multiple challenges.

First, current humanoid robotics technology exhibits a certain degree of specialization: while significant progress has been made in motion control, capabilities in cognitive understanding, dexterous manipulation, and environmental adaptability still fall short of market demands. Some argue that robots today continue to face substantial challenges in stability, durability, and dexterity, and that embodied intelligence remains confined to the experimental stage and data‑collection phase—far from achieving the real-time perception‑feedback and rapid decision‑making required for practical applications in production and daily life.

Secondly, the AI large models, precision machinery, and drive‑control systems that underpin embodied intelligence—components corresponding to the “brain–cerebellum–limbs” architecture—have yet to achieve seamless integration, creating technological bottlenecks in cross‑disciplinary convergence. Industry stakeholders recommend establishing a national key R&D program to strengthen upstream technological safeguards and overcome these coordination challenges across the brain–cerebellum–limbs framework.

Thirdly, high-quality, cross‑ontology whole-body motion data acquisition remains a bottleneck. Head‑body‑integrated AI companies often sidestep the core technical hurdle of motion control, while small and medium‑sized body‑only firms face even greater challenges in catching up on this front.

Faced with these challenges, industry consensus is becoming increasingly clear: because the “cerebellum” is primarily responsible for implementing motor control, its technological roadmap is characterized by high barriers and significant complexity, making it difficult to achieve breakthroughs through rapid catch-up. Companies that have already built up relevant technical capabilities will gain a competitive edge. For embodied intelligence systems, the “brain” handles scene understanding and task‑sequence generation, while the “cerebellum” ensures high‑precision, high‑reliability control of the physical hardware. No matter how “intelligent” an agent may be, bringing it into the real world still requires properly driving its “body.”

At present, humanoid robots are entering a new phase—the inaugural year of mass production—with over ten thousand units already being deployed in factory settings for training. Factory‑based training is pivotal to breaking the “gravitational constraint” between insufficient embodied‑intelligence data and limited practical applicability, and it is also an essential approach for accumulating high‑quality, large‑scale application data. Looking ahead, as force control and motion control become increasingly integrated, the “cerebellum” of embodied‑intelligence robots will continue to evolve, driving industrial robots steadily forward from mere “programmatic executors” toward true intelligent agents capable of closed‑loop perception, decision‑making, and execution.