The rapid evolution of generative artificial intelligence (AI) has ushered in a new era of disruptive innovation across industries. The advent of large language models (LLMs) like ChatGPT and Claude, followed by the rise of domestically developed counterparts such as DeepSeek, has demonstrated profound capabilities in semantic understanding and content generation. This progress has catalyzed the swift advancement of embodied AI, where physical forms interact with the environment, positioning it as a critical frontier for downstream AI applications. In this context, the humanoid robot, serving as a quintessential embodiment platform, has garnered unprecedented strategic importance. The Chinese Ministry of Industry and Information Technology’s 2023 “Guiding Opinions on the Innovative Development of Humanoid Robots” explicitly outlined goals to establish a preliminary innovation system and achieve breakthroughs in key technologies related to the “brain, cerebellum, and limbs” by 2025. Consequently, research and development (R&D) in humanoid robot technology has entered a phase of explosive growth.
While the field is burgeoning with new outputs, existing research often focuses on ethics, development directions, or specific applications. A systematic analysis of patent technological themes and their evolutionary pathways within China’s humanoid robot landscape remains underexplored. This gap limits a comprehensive understanding of the focal points and trajectory of technological advancement. To address this, I conducted an in-depth analysis using the BERTopic model on a corpus of 1,476 humanoid robot patent documents (titles and abstracts) from the PatSnap database. The study aimed to mine the thematic structure and reveal the dynamic evolution of these themes over time, providing valuable insights for strategic planning and industrial development in the humanoid robot domain.
Methodological Framework: BERTopic for Patent Analysis
My research methodology is designed to extract meaningful themes from unstructured patent text and trace their evolution. The process consists of three main phases: data collection and preprocessing, topic modeling, and evolutionary pathway analysis, as illustrated in the technical roadmap above.
1. Data Collection and Preprocessing: I retrieved patent documents from the PatSnap platform using a comprehensive query: TA:(“人形机器人” OR “拟人机器人” OR “类人机器人” OR “仿生人” OR “仿人机器人”). After data cleaning, 1,476 patents were obtained. The title and abstract of each patent were concatenated to form the core text for analysis, as abstracts concisely capture the technical problem, solution, and innovation. The publication year was extracted for temporal analysis. The annual application trend shows steady growth with a significant surge post-2023, aligning with policy directives and AI advancements.
| Step | Description |
|---|---|
| Tokenization | Segmenting Chinese text into meaningful words using Jieba. |
| Stop Word Removal | Eliminating common, non-informative words (e.g., “的”, “是”). |
| Custom Dictionary | Adding domain-specific terms (e.g., “步态”, “致动器”) to improve segmentation accuracy. |
| Text Cleaning | Removing punctuation, numbers, and other irrelevant characters. |
2. Topic Modeling with BERTopic: I employed the BERTopic model, which excels at capturing semantic nuances. The process involves:
- Semantic Embedding: Each preprocessed document was converted into a dense vector representation using the `Chinese-BERT-wwm` model, which understands contextual word meanings.
- Dimensionality Reduction: The high-dimensional embeddings were reduced using UMAP (Uniform Manifold Approximation and Projection) with parameters (n_neighbors=13, n_components=5, min_dist=0.1, metric=’cosine’) to preserve meaningful global structure while reducing noise.
- Clustering: The reduced embeddings were clustered using HDBSCAN (Hierarchical Density-Based Spatial Clustering) with parameters (min_cluster_size=13, min_samples=5, metric=’euclidean’) to identify dense regions of similar documents, forming potential topics. HDBSCAN effectively handles outliers.
- Topic Representation: For each cluster, the most representative words were extracted using a class-based TF-IDF (c-TF-IDF) procedure. This method calculates the importance of a word to a topic by comparing its frequency within the topic against its frequency across all documents. The formula for the weight $w_{t,c}$ of term $t$ in class (topic) $c$ is:
$$w_{t,c} = tf_{t,c} \times \log\left(\frac{N}{df_t}\right)$$
where $tf_{t,c}$ is the frequency of term $t$ in class $c$, $N$ is the total number of classes, and $df_t$ is the number of classes containing term $t$.
3. Evolutionary Pathway Analysis: To understand temporal dynamics, I divided the patent dataset into three distinct phases based on application volume and key technological milestones (e.g., the “Generative AI Year” of 2023). For each phase, a separate BERTopic model was built. The evolution between phases was mapped by calculating the cosine similarity between the top feature words of topics from consecutive phases. A threshold of 0.85 was set to identify significant evolutionary links, visualized using a Sankey diagram.
| Phase | Time Period | Patent Count | Characterization |
|---|---|---|---|
| P1: Germination | 2006–2014 | 274 | Early-stage exploration, low annual output. |
| P2: Accumulation | 2015–2022 | 724 | Steady growth, technology diversification. |
| P3: Growth | 2023–Present | 478 | Explosive growth, driven by AI/LLM integration. |
Holistic Snapshot: Six Pillars of Humanoid Robot Technology
Applying the BERTopic model to the entire corpus revealed six coherent technological themes, providing a macro-level view of China’s humanoid robot innovation landscape. The c-TF-IDF weights highlight the most salient keywords for each theme.
| Topic ID | Key Feature Words (with c-TF-IDF weights) | Interpreted Theme | Corresponding R&D Dimension |
|---|---|---|---|
| Topic 0 | 驱动 (0.121), 连接 (0.098), 关节 (0.095), 电机 (0.089), 组件 (0.085) | Joint Actuation & Mechanical Structure | Limb (Hardware Foundation) |
| Topic 1 | 规划 (0.132), 模型 (0.118), 步态 (0.115), 运动 (0.102), 轨迹 (0.094) | Gait Planning & Motion Control | Cerebellum (Locomotion & Stability) |
| Topic 2 | 关节 (0.125), 指骨 (0.112), 致动器 (0.108), 移动 (0.091), 操纵 (0.087) | Dexterous Hand & Finger Actuation | Limb (End-Effector) |
| Topic 3 | 数据 (0.141), 充电 (0.119), 信息 (0.106), 运动 (0.099), 采集 (0.092) | Data Acquisition & Power Management | Enabling Infrastructure |
| Topic 4 | 连接 (0.110), 动作 (0.105), 视觉 (0.101), 智能 (0.096), 头部 (0.089) | Visual Perception & Intelligent Decision-Making | Brain (Perception & AI) |
| Topic 5 | 制备 (0.136), 材料 (0.128), 电极 (0.115), 纤维 (0.104), 生物 (0.098) | Novel Material Fabrication & Biomimetics | Limb (Advanced Materials) |
The thematic structure aligns remarkably well with the conceptual framework of “Brain, Cerebellum, and Limb.” The Brain dimension is represented by Topic 4 (Visual Perception & Intelligent Decision-Making), focusing on multi-modal sensing and AI-driven decision-making empowered by models. The Cerebellum dimension is captured by Topic 1 (Gait Planning & Motion Control), dealing with stability, locomotion, and fine motor skills. The Limb dimension is prominently covered by Topics 0, 2, and 5, addressing the hardware trinity of joint/actuator mechanisms (Topic 0), dexterous end-effectors (Topic 2), and advanced materials (Topic 5). Topic 3 represents cross-cutting enabling technologies. This comprehensive coverage indicates a well-rounded and minimally gapped technological portfolio in China’s humanoid robot patent ecosystem.
Temporal Evolution: From Mechanical Foundations to Embodied Intelligence
A phase-wise analysis uncovers the dynamic trajectory of technological focus, highlighting a clear shift from foundational mechanics to integrated intelligence.
Phase 1 (P1): Germination (2006–2014) – Laying the Mechanical Groundwork
In this initial phase, patents concentrated on fundamental challenges. Two primary themes emerged:
- P1-Topic 0: Limb Support Structure & Basic Motion Control. Keywords: “支撑” (support), “动作” (action), “角度” (angle), “传感器” (sensor), “控制器” (controller). The focus was on achieving basic bodily support and simple motion control, exploring the fundamental hardware and control loops necessary for a humanoid robot to maintain posture and perform elementary movements.
- P1-Topic 1: Finger Joint Structure Design. Keywords: “指段” (finger segment), “适应” (adapt), “中部指段” (middle finger segment), “滑块” (slider). This theme underscores the early recognition of the anthropomorphic hand’s importance for interaction in human-centric environments. Research explored mechanical designs for finger joints to enhance adaptability and dexterity.
The P1 stage was characterized by exploratory work on core mechanical architectures and elementary control systems.
Phase 2 (P2): Accumulation (2015–2022) – Diversification and Specialization
This period saw significant growth and thematic branching into three distinct areas:
- P2-Topic 0: Drive Systems & Motion Models. Keywords: “驱动” (drive), “实现” (implement), “模型” (model), “转动” (rotate), “电机” (motor). Research deepened into the design of actuation systems and the mathematical modeling of their movement, indicating a move towards more sophisticated and validated control approaches.
- P2-Topic 1: Foot Mechanism & Gait Structure. Keywords: “脚趾” (toe), “机械” (mechanical), “装设” (install), “弹簧” (spring), “电机” (motor). Attention expanded to lower-limb mechanics, specifically foot design with elements like toes and springs to improve gait adaptability and energy efficiency.
- P2-Topic 2: Biomimetic Materials & Advanced Manufacturing. Keywords: “材料” (material), “生物” (biological), “尺寸” (dimension), “打印” (print), “韧带” (ligament). A novel theme emerged, integrating materials science. Keywords like “print” (3D printing) and “biological”/”ligament” point to the use of advanced fabrication and biomimicry to create lighter, more lifelike, and functional components.
P2 marked a transition from general structure to specialized subsystems and the introduction of new material paradigms.
Phase 3 (P3): Growth (2023–Present) – The Era of Intelligence and Integration
The current phase, likely propelled by generative AI breakthroughs, exhibits a dramatic thematic expansion into eight topics, which can be categorized into two overarching groups:
| Cluster | Topic ID & Feature Words | Core Focus |
|---|---|---|
| Hardware & Mechanical Refinement | P3-T0: 组件 (component), 大腿 (thigh), 小腿 (calf), 腿部 (leg), 转动 (rotate) | Lower-limb component design |
| P3-T1: 电机 (motor), 定子 (stator), 齿轮 (gear), 固定 (fix), 转子 (rotor) | Precision motor and transmission design | |
| P3-T6: 头部 (head), 收容 (house), 外壳 (shell), 表面 (surface), 组件 (component) | Head casing and assembly | |
| P3-T7: 指节 (knuckle), 手指 (finger), 手掌 (palm), 连杆 (link), 转动 (rotate) | Advanced hand joint kinematics (evolution of P1-T1) | |
| Intelligent Systems & Perception | P3-T2: 信息 (information), 图像 (image), 路径 (path), 场景 (scene), 点云 (point cloud) | Environmental perception & 3D scene understanding |
| P3-T3: 模型 (model), 状态 (state), 轨迹 (trajectory), 力矩 (torque), 动力学 (dynamics) | Dynamic trajectory planning & modeling | |
| P3-T4: 动作 (action), 训练 (training), 数据 (data), 学习 (learning), 位姿 (pose) | Data-driven motion learning & control | |
| P3-T5: 感知 (perception), 单元 (unit), 电极 (electrode), 压力 (pressure), 缓冲组件 (buffer) | Tactile sensing & sensor calibration |
The explosion of themes in the Intelligent Systems cluster is particularly telling. Keywords like “image,” “point cloud,” “training,” “learning,” and “model” directly reflect the integration of AI and large models (LMs) as the “brain” of the humanoid robot. The focus has shifted from merely controlling movement to enabling the robot to perceive, understand, plan, and learn from complex environments—closing the perception-decision-action loop. This signifies that the humanoid robot is evolving from a pre-programmed machine into an adaptable, learning-enabled embodied agent.
Mapping the Evolutionary Pathways
Calculating inter-phase topic similarities reveals a coherent technological lineage. The similarity between two topics $A$ and $B$ is computed based on their top-$k$ feature word vectors ($\mathbf{W_A}$, $\mathbf{W_B}$) derived from c-TF-IDF:
$$\text{Similarity}(A, B) = \frac{\sum_{i=1}^{k} w_{A,i} \cdot w_{B,i}}{\sqrt{\sum_{i=1}^{k} w_{A,i}^2} \cdot \sqrt{\sum_{i=1}^{k} w_{B,i}^2}}$$
where $w_{A,i}$ is the c-TF-IDF weight of the $i$-th word in topic $A$’s top-$k$ list.
The Sankey diagram generated from this analysis shows:
- P1 to P2: Foundational themes (P1-T0, P1-T1) branched out and fed into multiple P2 themes (Drive Systems, Foot Mechanics, Materials), demonstrating how early mechanical research enabled later specialization.
- P2 to P3: This is where convergence and intelligence infusion occur. The three P2 themes collectively gave rise to almost all P3 themes. For instance, P2-T0 (Drive Systems) evolved not only into refined motor design (P3-T1) but also contributed to intelligent control themes like trajectory planning (P3-T3) and motion learning (P3-T4). This illustrates how robust hardware platforms became the necessary substrate for advanced AI-driven functionalities.
The evolutionary path clearly demonstrates a logical progression: Mechanical Foundations (P1) → Subsystem Specialization (P2) → Intelligent, Modular Integration (P3).
Conclusions and Strategic Implications
My analysis of China’s humanoid robot patent landscape leads to several key conclusions and strategic recommendations.
Conclusions:
1. The holistic thematic structure is comprehensive, aligning with the “Brain, Cerebellum, Limb” framework and showing no major technological gaps at a macro level.
2. The technology has evolved along a clear trajectory: from initial focus on basic mechanical structure and control (P1), through a phase of diversified subsystem and material innovation (P2), to the current era of intelligent, AI-driven integration (P3).
3. The most significant shift, particularly post-2023, is the pervasive integration of AI/LLM-related technologies, transforming the humanoid robot from a mechanically sophisticated platform into an embodied intelligent agent capable of perception, learning, and decision-making in unstructured environments.
Strategic Recommendations:
| Focus Area | Challenge/Opportunity | Recommended Action |
|---|---|---|
| Deepen Technological Innovation | Real-time multi-modal sensor fusion and low-latency decision-making remain challenging. The “brain” layer, especially core algorithms and chips, shows a competitive gap compared to global leaders. | Accelerate R&D in synergistic hardware-software integration, focusing on proprietary solutions for high-performance, low-power compute architectures and real-time AI models tailored for embodied systems. Invest in core component independence (e.g., precision actuators, AI chips). |
| Explore Scenario Applications | P3 themes enable operation in complex environments, but a lack of commercial deployment and value闭环 (closed-loop value) persists. | Foster strong industry-academia collaboration for pilot deployments in targeted sectors (e.g., advanced manufacturing, elderly care, domestic services). Encourage R&D that is tightly coupled with real-world use cases to drive iterative improvement and validate economic models. |
| Perfect Standard Systems | The proliferation of specialized components and intelligent modules (as seen in P3’s eight themes) creates interoperability and safety challenges. | Establish a multi-layered standard system (national, industry, consortium). Prioritize standards for communication interfaces, safety protocols, performance evaluation, and data formats to ensure compatibility, accelerate industry collaboration, and guide safe, orderly development. |
In summary, the journey of the humanoid robot in China, as traced through its patent literature, is one of steady maturation from mechanics to mind. By understanding this evolutionary pathway and addressing the identified strategic imperatives, stakeholders can better navigate the exciting and complex landscape of building the next generation of embodied intelligent machines.