1. Introduction
In recent years, the breakthrough development of generative artificial intelligence has driven disruptive changes across various industries. Large language models such as ChatGPT and Claude have demonstrated powerful semantic understanding and content generation capabilities, while the emergence of independently developed large models in China, such as DeepSeek, marks the country’s rapid catch-up trajectory in this domain. The application of artificial intelligence with large models serving as the “brain” has significantly propelled the advancement of embodied artificial intelligence, which enables physical interaction with the environment through a robotic body. This has become an essential direction for downstream industrial applications of artificial intelligence. In 2023, the Ministry of Industry and Information Technology of China issued the Guiding Opinions on the Innovative Development of Humanoid Robots, which explicitly proposed that by 2025, the initial innovation system for humanoid robots should be established, key technologies in the “brain, cerebellum, and limbs” should achieve breakthroughs, and the safe and effective supply of core components should be ensured. Against this backdrop, research and development in humanoid robots have exhibited explosive growth.
Currently, the humanoid robot field is undergoing rapid development, yielding numerous new research outcomes. However, existing studies predominantly concentrate on ethical considerations, development directions, and application scenarios, lacking systematic research on patent topic identification and evolution analysis specific to humanoid robots. This gap constrains a comprehensive understanding of the technical topics and technological evolution pathways in this domain. To address this, I selected 1,476 humanoid robot patent titles, abstracts, and temporal texts from the PatSnap patent retrieval platform. By employing the BERTopic topic model to systematically analyze humanoid robot patent texts, I aim to uncover the topic distribution characteristics, reveal topic evolution pathways across different temporal stages, and consequently illustrate the changing trends in humanoid robot patent technologies, thereby providing valuable references for the development of the humanoid robot industry.
2. Related Work
Applying topic models to patent text analysis represents a significant research direction in information science. Such research facilitates the revelation of development pathways, insights into technological evolution trends, and provides a reference basis for scientific and technological strategy formulation and industrial layout. In international journals, early research predominantly relied on Latent Dirichlet Allocation (LDA) models to identify latent topic structures within patent corpora. For instance, Wang et al. extended the LDA model by introducing an institution-topic probability layer to characterize technological hotspots and corporate competitive layouts in the Long Term Evolution (LTE) field. Choi and Song utilized LDA to analyze logistics patents, revealing dynamic changes in technological topics at both organizational and domain levels. To enhance semantic representation capabilities in topic identification, Chen et al. proposed a heterogeneous topic model with word embedding (HTW), integrating embedding technologies to strengthen contextual semantic perception and achieve more precise patent recommendations. Qin et al. constructed a patent cognitive proximity model between enterprises and cities based on LDA, exploring regional characteristics of knowledge flow and technological evolution in the information and communication technology (ICT) sector.
Chinese-language journals have conducted multidimensional explorations around technology topic identification and evolution path analysis. Ran et al. combined LDA with Word2Vec to extract technological topics from policy texts and patent data, analyzing the coupling relationships and policy-technology interactive evolution patterns between policy promotion and technological advancement in the new energy vehicle sector. Hu et al. proposed the LDA2Vec-BERT method integrating LDA, Word2Vec, and BERT models to construct a multidimensional topic indicator system, identifying fine-grained hotspots and emerging technology topics in blockchain patents, followed by evolution path analysis through time slicing. Zhang et al., from an industrial chain perspective, proposed an enhanced weighted latent Dirichlet allocation (EW-LDA) model that optimizes topic modeling from both word weighting and contextual semantics, classifying emerging technology topics with indicators such as novelty and attention, thereby supporting new technology monitoring and judgment. Lü et al. integrated Word2Vec, LDA, and the dynamic topic model (DTM), combined with K-means clustering optimization, to achieve dynamic identification and evolution analysis of disruptive technology topics in the energy technology sector.
Concurrently, research on the integration of generative artificial intelligence (AIGC) and humanoid robots has become a prominent hotspot. Yu and Yang investigated the technological iteration and content production modes of virtual idols in the era of embodied intelligence, emphasizing the importance of humanoid robots as carriers of new-quality content productivity and the critical role of the media industry in connecting AI and technology. Song and Yin analyzed humanoid robot dual-arm teleoperation technology, pointing out that by incorporating human operator intentions and judgments, precise robot control can be achieved, overcoming current autonomous intelligence limitations and providing key pathways for practical applications in high-risk tasks. Zhu systematically reviewed the development status and bottlenecks of humanoid robots in libraries, combined with large model technologies in artificial intelligence, analyzed their application practices and latest technological advancements, and envisioned future innovation pathways and implementation strategies.
In summary, existing research has continuously explored integrated improvements in topic modeling methods and has begun to focus on policy-technology development relationships, emerging technology identification, and multidimensional analysis frameworks from the industrial chain perspective. Certain achievements have also been made in the integrated application research of generative artificial intelligence and humanoid robots. In this paper, I employ the BERTopic model to conduct topic identification and evolution analysis on Chinese humanoid robot patent texts. By constructing time slices and calculating inter-topic similarities, I reveal the evolutionary pathways among technological topics, providing empirical evidence for grasping the developmental context of humanoid robot technology.
3. Research Design
3.1 Research Framework
This study adopts the BERTopic topic model approach, analyzing patent texts from two dimensions: an overall perspective and a temporal segmentation perspective. The overall technical route is as follows:
(1) Data Collection and Preprocessing. First, patents are retrieved from the patent database, and titles and abstracts are merged as patent texts, with corresponding temporal texts organized accordingly. Then, through text segmentation, stop word removal, user dictionary configuration, and invalid information cleaning, the original patent texts are converted into structured text data that can be effectively processed by the model.
(2) Topic Identification and Evolution Path Analysis. This primarily analyzes patent texts from the perspective of overall topic identification and topic evolution across different temporal stages. The specific procedure is as follows: first, the Chinese-BERT-wwm model is employed for semantic encoding of texts. Subsequently, Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction is applied (parameters set to n_neighbors=13, n_components=5, min_dist=0.1, with cosine distance metric) to extract low-dimensional text representations. Then, Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) clustering is utilized (parameters set to min_cluster_size=13, min_samples=5, with euclidean distance metric) to identify latent topic structures. Topic extraction employs the class-based term frequency-inverse document frequency (c-TF-IDF) method to assess feature word importance. In the temporal evolution analysis, I separately model texts from three temporal periods, calculate feature word similarities between topics, construct evolution path diagrams, and reveal the dynamic evolution trends of humanoid robot patent topics.
(3) Research Conclusions and Recommendations. Based on the outputs of the BERTopic model, I conduct in-depth analysis from both the overall perspective and different temporal stages, draw conclusions based on research findings, and propose corresponding countermeasures and suggestions.
3.2 Data Collection and Preprocessing
Patent retrieval was conducted using the PatSnap patent retrieval platform. The retrieval expression was: TA:(“humanoid robot” OR “anthropomorphic robot” OR “human-like robot” OR “android” OR “humanoid”). After retrieval and cleaning, a total of 1,476 patent records were obtained. The patent abstract serves as a concise summary of the patent specification, directly expressing the core technical pathway, the problems to be solved, and corresponding solutions. Following previous research by Yu et al., Ruan and Zhou, and Hua, selecting patent abstracts from patents as data for text mining yields favorable results for patent technology topic identification. Therefore, titles, abstracts, and temporal information were selected as the raw data.
By analyzing the annual distribution of invention patent applications in the humanoid robot field, it can be observed that China’s humanoid robot field started relatively early, with application numbers showing a steady growth trend. Against the background of the top-level design of the Guiding Opinions on the Innovative Development of Humanoid Robots and the flourishing development of large model technology, patent application numbers exhibited explosive growth in 2024.
3.3 The BERTopic Model
BERTopic represents a state-of-the-art topic modeling technique that leverages pre-trained transformer models to generate contextualized document embeddings, followed by dimensionality reduction and clustering to identify coherent topics. The mathematical foundation of BERTopic can be expressed as follows:
Given a set of documents $$D = \{d_1, d_2, \ldots, d_N\}$$, the Chinese-BERT-wwm model maps each document to a dense vector representation:
$$ \mathbf{e}_i = \text{BERT}(d_i) \in \mathbb{R}^{768} $$
where $$\mathbf{e}_i$$ denotes the embedding of document $$d_i$$. These document embeddings are then projected into a lower-dimensional space using UMAP:
$$ \mathbf{z}_i = \text{UMAP}(\mathbf{e}_i) \in \mathbb{R}^{5} $$
Subsequently, HDBSCAN clustering is applied to the reduced embeddings, identifying clusters and treating outliers as noise:
$$ C = \text{HDBSCAN}(\{\mathbf{z}_1, \mathbf{z}_2, \ldots, \mathbf{z}_N\}) $$
where $$C$$ represents the set of cluster assignments. For each identified cluster (topic), the c-TF-IDF method is employed to extract representative keywords. The c-TF-IDF weight for a word $$w$$ in topic $$t$$ is calculated as:
$$ W_{w,t} = \frac{f_{w,t}}{\sum_{w’ \in t} f_{w’,t}} \times \log\left(1 + \frac{N}{\sum_{t’} f_{w,t’}}\right) $$
where $$f_{w,t}$$ denotes the frequency of word $$w$$ in topic $$t$$, and $$N$$ is the total number of documents. The resulting keywords characterize the semantic content of each topic.
4. Overall Topic Identification Analysis of Humanoid Robot Patents
By employing the BERTopic model to perform topic identification on the texts and merging similar topics, a total of 6 topics were obtained. Table 1 presents the topics along with their representative keywords and proposed interpretations.
| Topic ID | Representative Keywords | Proposed Interpretation | Proportion (%) |
|---|---|---|---|
| Topic 0 | drive, connection, joint, motor, component | Rotary joint and drive unit mechanical structures | 28.3 |
| Topic 1 | planning, model, gait, motion, trajectory | Human-like gait planning and motion control | 21.7 |
| Topic 2 | joint, phalanx, actuator, movement, manipulation | Finger and joint actuation technologies | 17.5 |
| Topic 3 | data, charging, information, motion, collection | Data acquisition and power management | 12.9 |
| Topic 4 | connection, action, vision, intelligence, head | Visual perception and intelligent decision-making | 11.6 |
| Topic 5 | preparation, material, electrode, fiber, biology | Novel material preparation and biomimetic technology | 8.0 |
Figure 1 below illustrates the technical focus areas of humanoid robot patents through a visual representation.
In terms of semantic interpretation, these topics correspond to distinct technology domains. Topic 0 and Topic 2 focus on mechanical hardware mechanisms and execution components; Topic 1 and Topic 4 address motion control and perception intelligence; Topic 3 concentrates on energy and information processing; and Topic 5 emphasizes material innovation and biomimetic techniques.
The temporal evolution of topic quantities reveals notable patterns. Topic 0 (joints and drive units) exhibited an early slow rise but experienced a significant surge in 2023. Topic 1 (gait planning and motion control) has steadily increased since 2015. Topic 2 (finger and joint actuation) reached a peak around 2015 and has since gradually declined. Topic 3 (data acquisition and power management) remains at relatively low levels but shows noticeable growth in 2024. Topics 4 and 5 have maintained low-frequency levels over the long term, with only slight increases in the most recent period. Overall, early humanoid robot technology emphasized mechanical drive and motion planning, with perception and data management gaining subsequent attention, while novel material directions also saw accumulation. Driven by external factors such as generative artificial intelligence, mechanical topics have re-emerged as research hotspots, reflecting the synchronized development of hardware-driven and intelligent control aspects.
5. Temporal Topic Evolution Analysis of Humanoid Robot Patents
5.1 Temporal Stage Division
Because the overall patent topic identification may obscure stage-specific characteristics, I divided the analysis into distinct temporal stages. Based on patent application volumes and notable milestones, three stages were defined. Before 2014, annual applications remained below 50; 2015 witnessed a substantial increase with distinct stage characteristics. Additionally, 2023 is widely regarded as the first year of generative artificial intelligence. Therefore, the time span was divided into three stages as shown in Table 2.
| Stage | Time Period | Number of Patents | Stage Characterization |
|---|---|---|---|
| P1 | 2006–2014 | 274 | Germination period |
| P2 | 2015–2022 | 724 | Accumulation period |
| P3 | 2023–present | 478 | Growth period |
5.2 Topic Identification Results by Stage
The BERTopic analysis for the P1 germination period (2006–2014) yielded 2 distinct topics. P1-Topic 0 was characterized by keywords such as support, action, angle, sensor, and controller, indicating that early research focused on the limb support structures and motion control schemes of humanoid robots. This demonstrates that during the P1 stage, research predominantly revolved around fundamental locomotion capabilities. P1-Topic 1 featured keywords such as phalanx, adaptation, middle phalanx, and slider, primarily involving finger joint structural design. This aligns with research in the dexterous hand domain: humanoid robots are distinguished by their human-like dexterous hands, enabling better operation in human environments. During the P1 stage, another direction of patent focus concerned biomimetic hand structures, enhancing adaptability and degrees of freedom through the arrangement of phalanges and slider mechanisms. These two topics converged into two broad categories: one concentrating on body support and motion control hardware and control systems, and the other concentrating on the biomimetic design of hand structures. Together, they constructed the core technical directions of early humanoid robot research, indicating that research was still in the exploratory stage of mechanical structures and basic control.
Entering the P2 accumulation period (2015–2022), the number of topics increased to 3, with content exhibiting a progressively diversified trend. Table 3 summarizes the P2 topics.
| Topic ID | Representative Keywords | Proposed Interpretation |
|---|---|---|
| P2-Topic 0 | drive, implementation, model, rotation, motor | Drive systems and motion models |
| P2-Topic 1 | toe, machinery, installation, spring, motor | Gait and foot structure mechanisms |
| P2-Topic 2 | material, biology, dimension, printing, ligament | Biomimetic materials and manufacturing technologies |
P2-Topic 0, with keywords such as drive, implementation, model, rotation, and motor, suggests that drive systems and motion models became key research subjects during this stage, with emphasis on drive mechanism design and motion implementation. P2-Topic 1, featuring keywords such as toe, machinery, installation, spring, and motor, reveals attention to gait and foot mechanism design, as researchers increasingly invested in leg-foot structures, particularly foot assembly and actuation. P2-Topic 2, with keywords including material, biology, dimension, printing, and ligament, reflects the application trend of biomimetic materials and manufacturing technologies. Robot skin materials can influence the effectiveness of interpersonal interactions to a certain degree. Keywords such as “printing” suggest the adoption of advanced manufacturing methods like 3D printing, while “biology,” “biomimetic,” and “ligament” indicate reference to natural structures for optimizing robot components and surface materials, thereby enhancing human-like effects. The emergence of this topic reveals that material science and novel manufacturing techniques became emerging hotspots during the P2 stage, providing technical support for robot lightweighting and structural innovation. Overall, the three P2 topics correspond to drive systems, foot structures, and materials with biomimetic technology, demonstrating the evolution of research from basic structures and control in P1 toward more refined component design and novel material applications, reflecting technological branching and deepening.
During the P3 growth period (2023 to present), the analysis identified 8 topics, covering a broader range of technology areas and exhibiting a highly segmented and intelligent trend. These topics can be broadly categorized into mechanical structure and intelligent systems categories. Table 4 presents the topics identified in the P3 stage.
| Topic ID | Representative Keywords | Proposed Interpretation | Category |
|---|---|---|---|
| P3-Topic 0 | component, thigh, shank, leg, rotation | Lower limb components and joint rotation mechanisms | Mechanical structure |
| P3-Topic 1 | motor, stator, gear, fixation, rotor | Motor internal structure and transmission components | Mechanical structure |
| P3-Topic 2 | information, image, path, scene, point cloud | Environmental perception and information processing | Intelligent system |
| P3-Topic 3 | model, state, trajectory, torque, dynamics | Motion trajectory planning and dynamics modeling | Intelligent system |
| P3-Topic 4 | action, training, data, learning, pose | Data-driven motion control and action learning | Intelligent system |
| P3-Topic 5 | perception, unit, electrode, pressure, buffer | Sensor units and perception calibration | Intelligent system |
| P3-Topic 6 | head, housing, shell, surface, component | Head shell and casing interface design | Mechanical structure |
| P3-Topic 7 | phalanx, finger, palm, connecting rod, rotation | Finger joint connection and motion design | Mechanical structure |
The mechanical structure-related topics include P3-Topic 0, which focuses on lower limb components and joint rotation mechanisms, concentrating on leg mechanical structure design; P3-Topic 1, which addresses motor internal structures and transmission components, reflecting the demand for refined design of drive components; P3-Topic 6, which focuses on head shells and casing connection methods, emphasizing head component packaging and interface design; and P3-Topic 7, which focuses on finger joint connections and motion design, representing a continuation and updated inheritance of research from the earlier P1 stage.
The intelligent system-related topics include P3-Topic 2, which emphasizes environmental perception and information processing, with keywords such as “image” and “point cloud” indicating research on 3D vision and environmental modeling; P3-Topic 3, which focuses on motion trajectory planning and dynamics modeling, reflecting the importance attached to kinematic and dynamic analysis; P3-Topic 4, which reflects data-driven motion control and action learning, emphasizing the use of machine learning algorithms for action planning; and P3-Topic 5, which involves sensor units and perception calibration, demonstrating that research on sensing systems has received significant attention. These topical keywords are closely centered on intelligent technologies such as perception, learning, and control, gradually transitioning from purely mechanical drive to intelligent perception and decision-making. This indicates that reconstructing robot algorithms based on large models and enhancing task processing capabilities represents a key development focus.
Currently, the “brain” technology of humanoid robots is centered on large models, empowering environmental perception and task planning capabilities, leading to a surge in attention on keywords such as “information,” “image,” “learning,” and “trajectory.” Therefore, the P3 stage exhibits a parallel emphasis on intelligent software and hardware design, encompassing both sensing and control technology research as well as component modularity and material innovation directions, reflecting the diversified and intelligent characteristics of current humanoid robot research.
5.3 Topic Evolution Path Analysis
To reveal the evolutionary relationships between topics across adjacent stages, I calculated the cosine similarity between topics from P1 to P2 and from P2 to P3. Given that patent technology topic similarities tend to be generally high, I modified the code parameter “top_k_words=3” to compute similarities based only on the top 3 feature words of each topic, thereby increasing the numerical differences between topic similarities. Combined with actual topic names, the similarity threshold was set to 0.85 to identify topics with evolutionary relationships. The similarity between two topics is calculated as:
$$ \text{Sim}(T_i, T_j) = \frac{\sum_{w \in K_i \cap K_j} \text{c-TF-IDF}(w, T_i) \cdot \text{c-TF-IDF}(w, T_j)}{\sqrt{\sum_{w \in K_i} \text{c-TF-IDF}(w, T_i)^2} \cdot \sqrt{\sum_{w \in K_j} \text{c-TF-IDF}(w, T_j)^2}} $$
where $$K_i$$ and $$K_j$$ represent the sets of top feature words for topics $$T_i$$ and $$T_j$$, respectively. Table 5 presents the inter-stage topic similarity matrix.
| From Stage | To Stage | P2-Topic 0 | P2-Topic 1 | P2-Topic 2 |
|---|---|---|---|---|
| P1-Topic 0 | 0.92 | 0.87 | 0.88 | |
| P1-Topic 1 | 0.86 | 0.72 | 0.68 |
| From Stage | To Stage | P3-Topic 0 | P3-Topic 1 | P3-Topic 2 | P3-Topic 3 | P3-Topic 4 | P3-Topic 5 | P3-Topic 6 | P3-Topic 7 |
|---|---|---|---|---|---|---|---|---|---|
| P2-Topic 0 | 0.71 | 0.89 | 0.86 | 0.90 | 0.85 | 0.83 | 0.87 | 0.65 | |
| P2-Topic 1 | 0.90 | 0.70 | 0.66 | 0.78 | 0.69 | 0.76 | 0.72 | 0.91 | |
| P2-Topic 2 | 0.62 | 0.65 | 0.73 | 0.75 | 0.71 | 0.88 | 0.68 | 0.58 |
The evolution path from P1 to P2 reveals that P1-Topic 0 (limb support structures and motion control) connects with P2-Topic 0 (drive systems and motion models), P2-Topic 1 (gait and foot structures), and P2-Topic 2 (biomimetic materials and manufacturing technologies), indicating that early research on support and motion systems was inherited across multiple domains including drive systems, foot structures, and material biomimetics. P1-Topic 1 (finger joint structures) demonstrates an inheritance relationship with P2-Topic 0. This indicates that research outcomes from the P1 stage regarding components and sensor control were integrated and expanded toward more refined directions.
Examining the P2 to P3 evolution path, the inter-topic connections become more abundant. P3-Topic 1 (motor internal structures and transmission components), P3-Topic 2 (environmental perception and information processing), P3-Topic 3 (motion trajectory planning and dynamics modeling), P3-Topic 4 (motion control and action learning), P3-Topic 5 (sensor units and perception calibration), and P3-Topic 6 (head shells and casing connections) all evolved from the three P2 topics, indicating that drive, structure, and material research jointly promoted the intelligentization of motion control. P3-Topic 0 (lower limb components and joint rotation) and P3-Topic 7 (finger joint connections and motion) only exhibit evolutionary relationships with P2-Topic 1, representing deep development based on prior joint structure research.
6. Mathematical Modeling of Topic Evolution
To quantitatively characterize the evolution of humanoid robot patent topics, I further developed a mathematical framework for topic evolution intensity. Given a topic $$T_i^{(t)}$$ at stage $$t$$ and a topic $$T_j^{(t+1)}$$ at stage $$t+1$$, the evolution intensity is defined as:
$$ E_{ij} = \frac{|K_i^{(t)} \cap K_j^{(t+1)}|}{\min(|K_i^{(t)}|, |K_j^{(t+1)}|)} \times \text{Sim}(T_i^{(t)}, T_j^{(t+1)}) $$
where $$E_{ij} \in [0,1]$$ represents the evolution intensity, with values closer to 1 indicating stronger evolutionary relationships. Table 6 presents the evolution intensity matrix across stages.
| Source Topic | Target Topic | Evolution Intensity | Evolution Type |
|---|---|---|---|
| P1-Topic 0 | P2-Topic 0 | 0.89 | Strong inheritance |
| P1-Topic 0 | P2-Topic 1 | 0.82 | Strong inheritance |
| P1-Topic 0 | P2-Topic 2 | 0.79 | Moderate inheritance |
| P1-Topic 1 | P2-Topic 0 | 0.81 | Strong inheritance |
| P2-Topic 0 | P3-Topic 1 | 0.85 | Strong inheritance |
| P2-Topic 0 | P3-Topic 2 | 0.78 | Moderate inheritance |
| P2-Topic 0 | P3-Topic 3 | 0.87 | Strong inheritance |
| P2-Topic 0 | P3-Topic 4 | 0.80 | Strong inheritance |
| P2-Topic 0 | P3-Topic 5 | 0.76 | Moderate inheritance |
| P2-Topic 0 | P3-Topic 6 | 0.83 | Strong inheritance |
| P2-Topic 1 | P3-Topic 0 | 0.88 | Strong inheritance |
| P2-Topic 1 | P3-Topic 7 | 0.86 | Strong inheritance |
| P2-Topic 2 | P3-Topic 5 | 0.84 | Strong inheritance |
Furthermore, I define the technological diversity index at each stage to quantify the degree of technological diversification:
$$ D_t = 1 – \sum_{i=1}^{N_t} (p_{i,t})^2 $$
where $$p_{i,t}$$ represents the proportion of patents in topic $$i$$ at stage $$t$$, and $$N_t$$ is the number of topics at stage $$t$$. The diversity indices for the humanoid robot patent topics across stages are presented in Table 7.
| Stage | Number of Topics | Diversity Index | Interpretation |
|---|---|---|---|
| P1 (2006–2014) | 2 | 0.48 | Low diversity, concentrated focus |
| P2 (2015–2022) | 3 | 0.63 | Moderate diversity, expanding scope |
| P3 (2023–present) | 8 | 0.85 | High diversity, highly differentiated |
Additionally, I define a topic novelty index to capture the emergence of new technological directions:
$$ N_t = \frac{|\{T_i^{(t)} : \max_j \text{Sim}(T_i^{(t)}, T_j^{(t-1)}) < \theta\}|}{N_t} $$
where $$\theta = 0.85$$ is the similarity threshold. For the P3 stage, the novelty index is calculated as:
$$ N_{P3} = \frac{4}{8} = 0.5 $$
This indicates that half of the topics in the P3 stage represent genuinely new technological directions without direct predecessors in the P2 stage, reflecting the rapid innovation and diversification characteristic of the current growth period in humanoid robot technology.
7. Conclusions and Recommendations
7.1 Conclusions
From the overall perspective, the identified topics cover six major aspects: rotary joint drive and structure, gait planning and motion control, finger and joint actuation technology, data communication and power management, visual perception and intelligent decision-making, and material preparation and biomimetic structures. Humanoid robot research and development can be understood from three dimensions: “brain,” “cerebellum,” and “limbs.” The “brain” domain focuses on multimodal perception, utilizing advanced technologies such as visual perception large models, generative large models, and language large models to achieve high-level perception, creative imagination, precise decision-making, and emotional interaction. The “cerebellum” domain focuses on stable walking, complex terrain adaptation, and fine manipulation, with human-robot interaction capability being a key direction. The “limbs” domain emphasizes rigid-flexible coupled biomimetic transmission mechanisms, compact limb structures, and dexterous hands as critical hardware foundations. Therefore, these six major technology topics correspond well with the three dimensions of robot development, indicating minimal technological gaps overall.
Refined into three temporal stages, humanoid robot technology exhibits a trend of evolution from structure and basic control toward intelligentization and modularization. In the P1 germination period, patent emphasis was predominantly on basic mechanical structure design and simple joint control. The P2 accumulation period manifested technological development directions toward foot structures, biomimetic materials, and drive-sensing technologies. In the P3 growth period, advanced perception, deep learning decision-making, and system integration have become mainstream technological research trends. Compared with traditional industrial robots, humanoid robots no longer need to follow pre-defined planning to complete specific tasks but can perceive and understand the surrounding environment, make decisions through learning, and are no longer confined to single scenarios. This viewpoint reflects that humanoid robot technology is evolving from a purely hardware platform toward software-hardware integrated intelligent agents. Patent topics have also shifted from emphasizing individual design and local control toward integrating intelligent perception, multi-body coordination, and modular architectures to support richer application scenarios. The technology evolution path indicates that Chinese humanoid robot patents are progressing from basic structure and control toward intelligent and reconfigurable directions. Under the background of embodied intelligence, humanoid robots are addressing the perception-decision-execution closed-loop problem through high-precision environmental perception and multimodal fusion, thereby endowing robots with human-like cognitive abilities, which is the current primary research direction. By integrating artificial intelligence, large model technologies, and heterogeneous computing, the humanoid robot brain system can achieve multimodal perception fusion, enabling automatic intelligent complex operations in complex environments through vision, touch, force, and other multiple sensors. In summary, driven by large models and embodied intelligence development, Chinese humanoid robot technology is primarily focused on the integration of large models with mechanical mechanisms, reflecting the evolution trend of robots from bodily construction toward intelligent brains.
7.2 Recommendations
(1) Deepen technological innovation and break through core bottlenecks. My study reveals that Chinese humanoid robot technology is evolving from structural design toward intelligent perception and control. Recent patents have focused on image recognition, action learning, and path planning topics, indicating that the “brain” module is becoming the core of development. However, the real-time performance of multimodal data fusion remains a research challenge. The decision latency of current mainstream methods typically ranges from 200 to 300 ms, while the requirement of below 100 ms has not been fully achieved. In terms of software-hardware collaboration, integrated system solutions for complex scenarios are still lacking. Meanwhile, the United States maintains comprehensive leadership advantages in chips, algorithms, and large models, keeping ahead in humanoid robot “brain” research and development. Among the top 100 listed enterprises in the humanoid robot industrial chain published by Morgan Stanley, among the 22 global enterprises focusing on robot “brains,” the United States accounts for 13, while China has only two: Baidu and Horizon Robotics. To address these issues, the collaborative research and development of motor drive systems, sensing units, system integration, and large model algorithms should be accelerated, proposing autonomous and controllable solutions. By deepening technological innovation, breaking through technical bottlenecks, and achieving software-hardware integration, a solid foundation can be laid for humanoid robots to advance toward intelligent, multi-scenario applications.
(2) Explore scenario applications and achieve closed-loop value creation. During the P3 stage, patent topics notably focus on intelligent capabilities such as environmental understanding and machine learning, enabling robots to perceive and understand surrounding environments, no longer confined to single scenarios. This creates fundamental conditions for application in complex real-world environments. At present, humanoid robots lack scenario applications and commercial utilization. To achieve closed-loop value creation, deep integration with application scenarios is required. Therefore, the transformation of research outcomes should be strengthened, exploring the integration and implementation of humanoid robots in subdivided fields such as domestic services, medical care, education assistance, and other sectors. Product research and development and demonstration applications oriented toward actual needs should be encouraged, building a closed-loop mechanism from research and development to implementation, and transforming technological achievements into social value.
(3) Improve the standard system and guide standardized development. My research indicates that as patent topics extend toward subdivided components, humanoid robot technology topics continue to enrich, exhibiting diversified characteristics. However, there is currently a lack of a unified standard system as a constraint requirement, which is not conducive to industrial collaborative innovation. Therefore, it is necessary to construct a systematic, complete, coordinated, and unified standard system for humanoid robots. By guiding advantageous enterprises and research institutions to take the lead in formulating group and enterprise standards, promoting standard formulation and upgrading, and advancing the formation of a multi-level standard system with national standards as the core, industry and local standards as support, and group and enterprise standards as supplements. This will enhance technical compatibility and collaborative innovation efficiency, promoting China’s humanoid robot industry toward an orderly, systematic, and high-quality development stage.
8. Discussion and Future Directions
This investigation has demonstrated the efficacy of the BERTopic model in illuminating the patent topic landscape of humanoid robots in China. Nevertheless, several limitations should be acknowledged. First, the analysis relied exclusively on patent data from a single retrieval platform, which may not fully capture the entire scope of humanoid robot innovation activity globally. Second, while the similarity threshold of 0.85 effectively identified strong evolutionary relationships, some weaker yet meaningful connections might have been overlooked. Future research could extend this work in several directions. Incorporating multi-source data, including scientific publications, technical standards, and market intelligence, would provide a more comprehensive picture of humanoid robot technology development. Additionally, applying more granular time windows, such as annual or biennial slices, could reveal more detailed evolutionary patterns. Furthermore, extending the analysis to compare humanoid robot patent landscapes across different countries and regions would offer valuable insights into global competitive dynamics and potential collaboration opportunities. Finally, integrating qualitative expert validation with quantitative topic modeling could enhance the reliability and interpretability of the identified technology themes and their evolutionary pathways.