A new BERTopic-based analysis of 1,476 humanoid robot patent records is offering a detailed view of how China’s humanoid robot technology landscape has developed and where it is heading. The study, conducted by researchers from the School of Public Administration at Xiangtan University, identifies six major humanoid robot technology themes across the full patent set and then divides the records into three time stages to trace how those themes have evolved. The central finding is that humanoid robot innovation in China has moved from an early emphasis on mechanical structures and basic control toward intelligent perception, data-driven motion learning and modular integration. The analysis does not rely on simple keyword counts alone. It combines Chinese-BERT-wwm text encoding, UMAP dimensionality reduction, HDBSCAN clustering and c-TF-IDF feature weighting to identify semantic patterns in humanoid robot patent titles and abstracts. The result is a structured map of humanoid robot technology that highlights both continuity and change.
The study’s patent sample covers records retrieved through a patent retrieval platform using search terms related to humanoid robot, anthropomorphic robot, human-like robot, bionic human and related humanoid robot variants. After cleaning, 1,476 records remained. Titles, abstracts and time information were used as the core text data. Patent abstracts were treated as especially useful because they summarize the technical route, the problem being addressed and the proposed solution. For a humanoid robot patent landscape, this matters because the abstract often reveals whether a patent is focused on a joint module, a gait-planning method, a hand actuator, a sensing unit, a power-management design, a material preparation process or an intelligent control model. The study therefore treats humanoid robot patent text as a semantic record of technical priorities rather than as a simple count of applications.
1. Patent text analysis builds a humanoid robot technology map
The research design has two layers. The first layer examines all humanoid robot patent records together to identify overall technology themes. The second layer divides the records into three time stages and models each stage separately. This two-layer approach is important because an overall humanoid robot patent model can reveal the broad structure of the field, while stage-specific models can show which humanoid robot themes are emerging, maturing or fading. The method uses Chinese-BERT-wwm to encode the semantic meaning of patent text. It then applies UMAP for dimensionality reduction with parameters set to 13 neighbors, 5 components, 0.1 minimum distance and cosine distance. HDBSCAN is used for clustering with a minimum cluster size of 13, 5 minimum samples and Euclidean distance. Topic features are extracted with c-TF-IDF, which measures how important a term is to a topic relative to other topics. For the stage analysis, the study calculates similarity between themes in adjacent periods and uses a threshold to identify likely evolution paths.
The time division is based on observable changes in humanoid robot patent applications. Before 2014, annual applications remained below 50. In 2015, applications rose sharply, marking a clear shift. The year 2023 is treated as the start of a new stage because it is widely associated with the rise of generative artificial intelligence, and the study specifically asks whether generative artificial intelligence and large-model development affected humanoid robot patent themes. The three stages are therefore Stage 1, the萌芽期 or early stage from 2006 to 2014 with 274 patents; Stage 2, the accumulation stage from 2015 to 2022 with 724 patents; and Stage 3, the growth stage from 2023 onward with 478 patents. Across the full set, humanoid robot patent applications show a stable early increase followed by a sharp rise in 2024, reflecting the combined influence of top-level policy design and rapid advances in large-model technology.
| Stage | Period | Humanoid robot patent records | Themes identified |
|---|---|---|---|
| Stage 1 | 2006–2014 | 274 | 2 |
| Stage 2 | 2015–2022 | 724 | 3 |
| Stage 3 | 2023–present | 478 | 8 |
The stage counts are not merely descriptive. They show how the humanoid robot patent field has become more differentiated. Stage 1 produced two broad themes, Stage 2 produced three themes and Stage 3 produced eight themes. For humanoid robot technology, this expansion suggests that the field is no longer organized around a small number of basic mechanical questions. Instead, humanoid robot patents now cover a wider set of technical problems, including lower-limb components, motor internals, environmental perception, trajectory planning, action learning, sensor calibration, head-shell design and finger-joint mechanisms. The rise in theme number is therefore a sign of technical subdivision, not simply a rise in patent volume.
2. Six overall humanoid robot themes cover core hardware, control, sensing and materials
When all 1,476 humanoid robot patent records are modeled together, the study identifies six major themes. These themes are broad, but they are technically meaningful. They cover joint and drive structures, gait planning and motion control, hand and finger actuation, data collection and power management, visual perception and intelligent decision-making, and material preparation and biomimetic structures. The presence of these six themes suggests that the overall humanoid robot patent landscape has relatively few obvious technology gaps at the broad thematic level. The field contains hardware, control, energy, perception and material strands, even if the balance among them changes over time.
| Overall humanoid robot theme | Representative feature terms | Humanoid robot technology focus |
|---|---|---|
| Topic 0 | drive, connection, joint, motor, component | Joint and drive-unit mechanical structures for the humanoid robot body |
| Topic 1 | planning, model, gait, motion, trajectory | Humanoid robot gait planning and motion control |
| Topic 2 | joint, phalanx, actuator, movement, manipulation | Hand, finger-joint and actuation technology for humanoid robot manipulation |
| Topic 3 | data, charging, information, motion, collection | Data collection, information handling and power management for humanoid robot operation |
| Topic 4 | connection, action, vision, intelligence, head | Visual perception, intelligent decision-making and head-related humanoid robot systems |
| Topic 5 | preparation, material, electrode, fiber, biology | New material preparation and biomimetic technology for humanoid robot components |
The first overall humanoid robot theme, Topic 0, centers on drive, connection, joint, motor and component. This is a hardware-oriented humanoid robot theme. It reflects the mechanical reality that a humanoid robot requires joints and drive units that can support movement, load and control. The second theme, Topic 1, focuses on planning, model, gait, motion and trajectory. This is the humanoid robot locomotion and control theme. It concerns how a humanoid robot plans steps, generates movement and follows trajectories. The third theme, Topic 2, centers on joint, phalanx, actuator, movement and manipulation. This is a humanoid robot hand and finger-actuation theme, and it reflects the importance of dexterous manipulation in humanoid robot design. The fourth theme, Topic 3, includes data, charging, information, motion and collection. This humanoid robot theme points to data acquisition, information flow and power management, which are essential for continuous operation and intelligent behavior.
The fifth overall humanoid robot theme, Topic 4, includes connection, action, vision, intelligence and head. This theme links visual perception, intelligent decision-making and head-related humanoid robot structures. It is a bridge between the mechanical humanoid robot body and the intelligent humanoid robot system. The sixth theme, Topic 5, includes preparation, material, electrode, fiber and biology. This humanoid robot theme highlights new materials, electrodes, fibers and biomimetic approaches. It shows that humanoid robot innovation is not limited to motors and algorithms. Material science and biologically inspired design are also part of the humanoid robot patent landscape. Taken together, the six overall humanoid robot themes align with the common three-part understanding of humanoid robot development: the brain, the cerebellum and the limbs. The brain dimension involves multimodal perception, visual models, generative models, language models, decision-making and interaction. The cerebellum dimension involves stable walking, complex terrain adaptation and fine operation. The limb dimension involves rigid-flexible coupling transmission, compact limb structures and dexterous hands. The six themes map onto these dimensions with relatively little thematic absence.
3. Time-stage modeling reveals a humanoid robot shift toward intelligence
The overall humanoid robot themes are useful, but they can hide change. The study therefore models each stage separately. Stage 1, from 2006 to 2014, contains 274 humanoid robot patent records and produces two themes. Stage 2, from 2015 to 2022, contains 724 records and produces three themes. Stage 3, from 2023 onward, contains 478 records and produces eight themes. This pattern shows a clear expansion in humanoid robot technical differentiation. In Stage 1, humanoid robot patents are organized around basic support structures, motion control and finger-joint design. In Stage 2, the humanoid robot field deepens into drive systems, foot and gait mechanisms, and biomimetic materials and manufacturing. In Stage 3, humanoid robot patents split into a wider array of mechanical and intelligent-system themes. The direction of travel is from structure and basic control toward intelligence, modularity and system integration.
The study also examines how humanoid robot themes change in volume over time. Topic 0, the joint and drive-unit theme, rises slowly in the early years and then shows a significant surge in 2023. Topic 1, the gait-planning and motion-control theme, grows gradually from 2015 and rises steadily in recent years. Topic 2, the finger and joint-actuation theme, peaks around 2015 and then gradually declines. Topic 3, the data collection and power-management theme, remains relatively low in volume but shows a clear increase in 2024. Topic 4, the visual perception and intelligent decision-making theme, and Topic 5, the new material and biomimetic theme, remain low in frequency for a long period and increase only slightly in the latest stage. These patterns suggest that early humanoid robot work concentrated on mechanical drive and motion planning, while later humanoid robot work began to pay more attention to perception, data management and new materials. With the external push from generative artificial intelligence, the mechanical humanoid robot theme becomes a hotspot again, showing that hardware drive and intelligent control are developing together rather than replacing each other.
4. Stage 1 humanoid robot patents focused on support structures and basic control
In Stage 1, the humanoid robot patent model identifies two themes. The first theme includes support, action, angle, sensor and controller. This humanoid robot theme points to limb support structures and motion-control solutions. Early humanoid robot research was concerned with whether the machine could support itself, move in a controlled way and use basic sensing and control to maintain action. This is consistent with a field still exploring the mechanical and control foundations of humanoid robot movement. The second Stage 1 theme includes phalanx, adaptation, middle phalanx and slider. This humanoid robot theme concerns finger-joint structure design. It reflects the importance of humanoid robot hands that can operate in human environments. The humanoid robot is distinguished partly by its human-like body and dexterous hand, and early patents explored phalanx and slider mechanisms to improve hand adaptability and degrees of freedom.
The two Stage 1 humanoid robot themes can be grouped into two early technical directions. One direction is the hardware and control system for body support and motion. The other is the biomimetic design of hand structures. Together, these themes define the early core of humanoid robot technology. They show that Stage 1 humanoid robot research was still largely at the level of mechanical structure and basic control exploration. The humanoid robot patent record in this period is not yet dominated by learning, perception or large-model integration. Instead, it is dominated by the physical conditions that make humanoid robot movement and manipulation possible.
5. Stage 2 humanoid robot patents deepen drive, foot and material subsystems
In Stage 2, the humanoid robot patent model identifies three themes, and the content becomes more diversified. The first Stage 2 theme includes drive, implementation, model, rotation and motor. This humanoid robot theme shows that drive systems and motion models became key research objects. The focus is on drive-mechanism design and motion realization, and the appearance of model-related terms suggests that researchers began to build and validate humanoid robot drive models. The second Stage 2 theme includes toe, mechanical, installation, spring and motor. This humanoid robot theme shows attention to gait and foot-mechanism design. In Stage 2, humanoid robot research began to invest more in leg and foot structures, especially foot assembly and drive. The third Stage 2 theme includes material, biology, size, printing and ligament. This humanoid robot theme reflects the application of biomimetic materials and manufacturing technologies. Terms such as printing point to advanced manufacturing methods such as 3D printing, while biology and ligament suggest reference to natural structures to optimize humanoid robot components and surface materials. Robot skin materials can influence human interaction, and biomimetic materials can support lightweight design and structural innovation.
Overall, the three Stage 2 humanoid robot themes correspond to drive systems, humanoid robot foot structures, and materials and biomimetic technology. The research focus moves from the basic structures and control of Stage 1 toward more detailed component design and new material application. This is a period of branching and deepening. The humanoid robot patent landscape becomes more specific. Instead of only asking how a humanoid robot body can be supported and controlled, Stage 2 patents ask how drive systems can be modeled, how feet and toes can be mechanically arranged, and how materials and manufacturing can make humanoid robot structures more adaptive. The humanoid robot field is beginning to specialize.
6. Stage 3 humanoid robot patents expand into eight mechanical and intelligent themes
In Stage 3, from 2023 onward, the humanoid robot patent model identifies eight themes. These themes cover a wider technical space and show a strong trend toward subdivision and intelligence. They can be roughly divided into mechanical-structure themes and intelligent-system themes. On the mechanical side, one humanoid robot theme includes components, thigh, calf, leg and rotation, focusing on lower-limb components and joint-rotation mechanisms. Another includes motor, stator, gear, fixed and rotor, focusing on motor internals and transmission parts. Another includes head, housing, shell, surface and component, focusing on head shells and housing interfaces. Another includes phalanx, finger, palm, linkage and rotation, focusing on finger-joint connection and motion design. These themes continue the earlier humanoid robot concern with hardware, but they are more detailed and more component-specific.
| Stage 3 humanoid robot theme | Representative feature terms | Humanoid robot technology focus |
|---|---|---|
| Topic 0 | component, thigh, calf, leg, rotation | Lower-limb components and joint-rotation mechanisms |
| Topic 1 | motor, stator, gear, fixed, rotor | Motor internals and transmission parts |
| Topic 2 | information, image, path, scene, point cloud | Environmental perception and information processing |
| Topic 3 | model, state, trajectory, torque, dynamics | Trajectory planning and dynamics modeling |
| Topic 4 | action, training, data, learning, pose | Data-driven motion control and action learning |
| Topic 5 | perception, unit, electrode, pressure, buffer component | Sensor units and perception calibration |
| Topic 6 | head, housing, shell, surface, component | Head shells and housing interface design |
| Topic 7 | phalanx, finger, palm, linkage, rotation | Finger-joint connection and motion design |
On the intelligent-system side, the humanoid robot themes are equally important. One theme includes information, image, path, scene and point cloud, pointing to environmental perception and information processing. Terms such as image and point cloud indicate three-dimensional vision and environment modeling. Another theme includes model, state, trajectory, torque and dynamics, pointing to motion-trajectory planning and dynamics modeling. Another includes action, training, data, learning and pose, reflecting data-based motion control and action learning. Another includes perception, unit, electrode, pressure and buffer component, indicating sensor units and perception calibration. These humanoid robot themes are closely tied to perception, learning and control. They show that humanoid robot technology is moving from purely mechanical drive toward intelligent perception and decision-making. The rise of large models is especially relevant because large models can enhance humanoid robot environment perception and task planning, which helps explain the increased attention to information, image, learning and trajectory in Stage 3 humanoid robot patents.
The Stage 3 humanoid robot patent landscape therefore shows a dual structure. Hardware design remains important, but it is increasingly accompanied by intelligent software and sensing. The humanoid robot is not only a mechanical platform. It is becoming a software-hardware integrated intelligent system. The eight Stage 3 humanoid robot themes reflect this: lower-limb structures, motor internals, environmental perception, trajectory planning, action learning, sensor calibration, head-shell design and finger-joint mechanisms all appear as distinct patent topics. This is a more modular and more intelligent humanoid robot technology landscape than the one seen in Stage 1 or Stage 2.
7. Evolution paths connect early humanoid robot structures to later intelligent control
The study calculates cosine similarity between themes in adjacent stages. Because patent technology themes often show generally high similarity, the analysis uses the top three feature words for each theme and sets a similarity threshold of 0.85 to identify meaningful evolution relationships. The resulting evolution map shows how humanoid robot themes connect across stages. From Stage 1 to Stage 2, the early humanoid robot theme on limb support structures and motion control connects to Stage 2 themes on drive systems and motion models, gait and foot structures, and biomimetic materials and manufacturing. This means that early humanoid robot work on support and motion was inherited and extended into several more specific directions. The Stage 1 humanoid robot theme on finger-joint structure also has a relationship with the Stage 2 drive-system and motion-model theme. Stage 1 humanoid robot component and sensor-control research was therefore integrated and expanded into more detailed directions. The early humanoid robot support, action and hand-structure themes provided a technical base for later drive design, spring-toe mechanisms and material applications. The evolution is from overall mechanical architecture toward specific components and material innovation.
From Stage 2 to Stage 3, the connections become richer. Stage 3 humanoid robot themes on motor internals and transmission parts, environmental perception and information processing, motion-trajectory planning and dynamics modeling, motion control and action learning, sensor units and perception calibration, and head shells and housing interfaces all evolve from the three Stage 2 humanoid robot themes. This indicates that drive, structure and material research, together with machine learning, jointly pushed forward the intelligentization of humanoid robot motion control. The Stage 3 humanoid robot themes on lower-limb components and joint rotation and on finger-joint connection and motion are linked mainly to the Stage 2 foot and gait theme, showing that earlier joint-structure research was deepened further. The evolution paths therefore reveal a logic in which early structure and drive research gradually moves toward multimodal perception and intelligent control.
The stage-based humanoid robot evolution path can be summarized as follows. Stage 1 laid the mechanical structure foundation. Stage 2 deepened drive, foot and material subsystems. Stage 3 produced many new humanoid robot themes related to perception and large models. Large models can significantly compensate for the shortage and specialization of training data in the robot field, providing humanoid robot systems with stronger perception, understanding, decision-making and action capabilities. The new Stage 3 humanoid robot themes also point to future expansion into modular components and new sensing materials. Overall, the evolution map shows both inheritance and transformation in humanoid robot technology. The mechanical and sensing technologies established in earlier stages became the foundation for later complex intelligent functions. Themes from different stages connect to one another and together shape the current diversified humanoid robot research landscape.
8. The humanoid robot field is moving from body construction to intelligent brains
The study’s conclusions can be understood through the three dimensions of brain, cerebellum and limbs. In the brain dimension, humanoid robot development focuses on multimodal perception, visual perception models, generative models and language models to achieve advanced perception, creative imagination, precise decision-making and emotional interaction. In the cerebellum dimension, humanoid robot development focuses on stable walking, complex terrain adaptation and fine operation, with human-machine interaction as a key direction. In the limb dimension, humanoid robot development focuses on rigid-flexible coupling biomimetic transmission mechanisms, compact limb structures and dexterous hands. The six overall humanoid robot themes match these three dimensions with relatively little technical absence. This is an important finding because it suggests that China’s humanoid robot patent landscape is not narrowly concentrated in one subsystem. It contains hardware, control, sensing, energy and material elements.
When the analysis is refined into three stages, the humanoid robot field shows a clear trend from structure and basic control toward intelligence and modularization. In Stage 1, humanoid robot patents focused on basic mechanical structure design and simple joint control. In Stage 2, humanoid robot patents moved toward foot structures, biomimetic materials and drive sensing. In Stage 3, advanced perception, deep-learning decision-making and system integration became mainstream humanoid robot research trends. Unlike traditional industrial robots, humanoid robots no longer need to follow pre-set plans to complete specific tasks. They are expected to perceive and understand their surroundings, learn from data and make decisions, and operate beyond a single scenario. This shows that humanoid robot technology is evolving from a pure hardware platform into a software-hardware integrated intelligent agent. Humanoid robot patent themes are shifting from single-unit design and local control toward integrated intelligent perception, multi-body coordination and modular architecture that can support richer application scenarios.
The humanoid robot evolution path shows that China’s humanoid robot patents are moving from basic structure and control toward intelligence and reconfigurability. In the context of embodied intelligence, a major research direction is to start with high-precision environmental perception and multimodal fusion, solve the perception-decision-execution loop and give the humanoid robot human-like cognitive capabilities. By integrating artificial intelligence, large-model technology and heterogeneous computing, the humanoid robot brain system can achieve multimodal perception fusion. It can use vision, touch, force and other sensors to perform automatic intelligent complex operations in complex environments. Under the push of large models and embodied intelligence, China’s humanoid robot technology is mainly focused on the integration of large models and mechanical mechanisms. This reflects a broader shift in the humanoid robot field from body construction toward intelligent brains.
9. Humanoid robot development faces bottlenecks in real-time fusion and integrated systems
The study identifies several practical bottlenecks for humanoid robot development. The first is real-time multimodal data fusion. The analysis shows that humanoid robot technology is moving from structural design toward intelligent perception and control, with recent humanoid robot themes focusing on image recognition, action learning and path planning. This makes the brain module a core development area. However, the real-time performance of multimodal data fusion remains a research difficulty. Mainstream methods often have decision latency of 200 to 300 milliseconds, and the requirement of below 100 milliseconds has not yet been fully achieved. In software-hardware coordination, there is still a lack of integrated system solutions for complex scenarios. The study also notes that the United States has comprehensive advantages in chips, algorithms and large models and maintains leadership in humanoid robot brain research. In a Morgan Stanley list of top 100 listed companies in the humanoid robot industry chain, 22 companies globally focus on the robot brain, including 13 in the United States and only two in China, Baidu and Horizon Robotics. These observations point to the need for coordinated research and development in motor drive, sensing units, systems and large-model algorithms, and for autonomous and controllable solutions.
The second bottleneck concerns application scenarios and value closure. In Stage 3, humanoid robot patent themes clearly focus on environmental understanding and machine learning, enabling the humanoid robot to perceive and understand its surroundings rather than remain limited to a single scenario. This creates a foundation for application in complex real environments. At present, however, humanoid robots lack scenario application and commercial use. To achieve value closure, humanoid robot development must be deeply integrated with application scenarios. The study recommends strengthening technology transfer and exploring humanoid robot deployment in household services, medical care, educational assistance and other segmented fields. It also recommends encouraging product development and demonstration applications oriented toward real needs and building a closed-loop mechanism from research and development to implementation so that humanoid robot technology can be transformed into social value.
The third bottleneck concerns standards. As humanoid robot patent themes extend to more detailed components, the humanoid robot technology landscape becomes richer and more diverse. However, there is currently a lack of a unified standards system to guide industry coordination. The study therefore recommends building a systematic, complete and coordinated humanoid robot standards system. Leading enterprises and research institutions should take the lead in formulating group and enterprise standards, promote standard formulation and upgrading, and form a multi-level standards system with national standards as the core, industry and local standards as support, and group and enterprise standards as supplements. This would improve technical compatibility and collaborative innovation efficiency and help China’s humanoid robot industry move toward an orderly, systematic and high-quality development stage.
10. Humanoid robot competition will depend on integrated innovation
The humanoid robot patent landscape described in the study is not a static list of technologies. It is a dynamic record of how the field has organized itself over time. The overall six themes show that humanoid robot technology already covers joints and drives, gait and motion control, hand actuation, data and power management, visual perception and intelligent decision-making, and materials and biomimetic structures. The three-stage analysis shows that humanoid robot innovation has moved from early support structures and finger-joint design to drive systems, foot mechanisms and biomimetic materials, and then to lower-limb components, motor internals, environmental perception, trajectory planning, action learning, sensor calibration, head-shell design and finger-joint mechanisms. The evolution paths connect these stages and show that earlier humanoid robot mechanical and sensing research became the foundation for later intelligent humanoid robot functions.
For the humanoid robot industry, the implication is that progress will depend on integrated innovation. A humanoid robot cannot advance through mechanics alone, and it cannot advance through algorithms alone. It requires joint modules, actuators, sensors, power systems, materials, perception models, motion-planning methods and action-learning systems to work together. The humanoid robot patent data show that China’s humanoid robot field is already broad, but the next phase will require deeper integration across hardware and software. The study’s recommendations point in this direction: deepen technological innovation to overcome core bottlenecks, explore scenario applications to achieve value closure, and improve standards systems to guide standardized development. These recommendations are directly connected to the patent evidence. The humanoid robot brain is becoming central, but the humanoid robot body remains essential. The humanoid robot field is moving toward intelligent, modular and integrated systems, and the patent landscape reflects that shift.
The broader message is that humanoid robot development is entering a phase in which technical breadth must be matched by technical depth. The six overall humanoid robot themes show broad coverage. The three-stage humanoid robot evolution shows a movement toward intelligence. The remaining challenges are real-time multimodal fusion, integrated software-hardware solutions, application validation and standards coordination. How effectively the humanoid robot field addresses these challenges will shape the next stage of humanoid robot patent activity and humanoid robot industrial development. The patent record suggests that China’s humanoid robot innovation has built a substantial mechanical and control base and is now pushing toward intelligent perception, learning and system integration. The humanoid robot of the future will be judged not only by its structure or its motor performance, but by its ability to perceive, decide and act in complex real-world environments. That is the direction the humanoid robot patent landscape is already pointing toward.