As we delve into the era of artificial intelligence, the humanoid robot stands out as a pivotal fusion of AI and the physical world, representing a critical frontier in the global industrial revolution. We believe that deeply laying out the humanoid robot industry is essential for seizing the future industrial high ground. This article systematically explores the strategic significance and pathways for developing the humanoid robot sector, drawing on technological evolution, industrial ecosystem construction, and international comparisons. We propose that breakthroughs in technology, ecosystem building, and policy coordination can enable a “lane-changing overtaking” in the universal intelligent tool revolution, establishing a global competitive edge. The humanoid robot, as a general-purpose intelligent tool, is poised to redefine productivity and become a cornerstone of future economic landscapes.
Historically, leaps in productivity have always been intertwined with technological and industrial revolutions, often manifesting through innovations in production tools. From steam engines to computers, each advancement has mechanized and automated processes, liberating human labor. Today, with the rapid development of AI and big data, new-generation tools extend human capabilities not just physically but intellectually, driving what we term new qualitative productive forces. The humanoid robot epitomizes this shift, serving as an optimal载体 for AI to interact with the physical environment. Unlike traditional robots reliant on rule-based algorithms, AI-driven humanoid robots exhibit higher intelligence and cross-task versatility, thanks to end-to-end neural network control methods. These methods enable direct mapping from sensor inputs to control commands, enhancing motion naturalness and adaptability. For instance, in motion control, neural networks learn from data like visual information to output precise instructions, improving autonomy. Consider the equation for an end-to-end neural network control: $$ \mathbf{u} = f(\mathbf{s}, \theta) $$ where \(\mathbf{u}\) represents the control output, \(\mathbf{s}\) is the sensor input vector, and \(\theta\) denotes the network parameters optimized through training. This approach allows humanoid robots to simulate human-like movements and make autonomous decisions, as seen in models like Tesla’s Optimus, which uses such networks for object classification and limb calibration.
The humanoid robot is set to become one of the most productive tools of the Fourth Industrial Revolution. Its human-like form, with dexterous hands, allows it to adapt to ergonomically designed environments without extensive modifications. As intelligence evolves from narrow AI toward general AI, humanoid robots can better understand human intent, execute tasks efficiently, and interact naturally. They function as multi-task, general-purpose platforms, expandable via sensors, actuators, or software modules for diverse scenarios like industrial manufacturing, commercial services, and home companionship. With global aging and rising labor costs, humanoid robots offer a solution to workforce shortages. We estimate that in the coming decades, the number of humanoid robots could surpass humans, potentially reaching over 10 billion units by 2040, underscoring their role as a strategic pillar industry and a variable reshaping the global economy.
To harness this potential, we must master a new industrial ecosystem centered on the humanoid robot, comprising three core parts: the本体, intelligent operating systems, and intelligent applications, along with non-core elements like testing and simulation. Dominance in these areas is crucial for high-quality development. The本体, or physical base, includes the frame, joints, actuators, transmission systems, end-effectors, sensor integration, power sources, control systems, communication interfaces, safety mechanisms, and environmental adaptability components. These determine the upper limit of application capabilities and represent the “last mile” for deployment. While core components such as reducers, motors, and sensors have seen significant advances in power density and cost, they still fall short of humanoid robot demands, particularly in areas like the “dexterous hand,” which faces challenges in sensitivity and joint matching. We advocate for intensified innovation to overcome these bottlenecks.
Intelligent operating systems are the heart of the humanoid robot ecosystem, differing from traditional ones by emphasizing smart adaptability that dynamically adjusts to complex environments. Just as iOS standardized mobile ecosystems, a standardized intelligent OS for humanoid robots can spur industry-wide collaboration and innovation. By developing systems like Open Harmony, we can avoid foreign dependency and ensure supply chain security. The core of such systems lies in their ability to generate autonomous processes, as modeled by: $$ \pi(a|s) = \text{softmax}(g(s, a)) $$ where \(\pi\) is the policy for action \(a\) given state \(s\), and \(g\) is a function learned through reinforcement and imitation learning. This enables humanoid robots to perform tasks like navigation and object manipulation with minimal human intervention.
Intelligent applications unlock the humanoid robot’s vast potential and ecological effects. In industry, they enhance safety in hazardous operations; in services, they integrate into daily life for tasks like housekeeping and medical assistance; and in special fields, they aid in rescue and exploration. Goldman Sachs predicts the humanoid robot market could hit $154 billion by 2035, rivaling electric vehicles. Moreover, the industry fosters aggregation effects, benefiting underlying technologies like AI models and simulation software, and spurring innovation in related sectors such as R&D and maintenance. The following table summarizes the current state of humanoid robot本体 production, highlighting domestic and international comparisons:
| Domain | Domestic Leaders | International Leaders | Domestic Strengths | Areas for Improvement |
|---|---|---|---|---|
| Reducers | LVD Harmonic, Laifu Harmonic, Qinchuan Machine Tool | Nabtesco (Japan), Sumitomo Heavy Industries (Japan) | Harmonic reducers | RV reducers, planetary reducers |
| Motors | Estun Automation, Siasun, Guangzhou CNC | FANUC (Japan), Yaskawa (Japan), KUKA (Germany) | Stepper motors, frameless torque motors | Hollow cup motors |
| Lead Screws | Luoyang Bearing, Wuzhou Xinchun, Dingzhi Tech | Rexroth (Germany), Thomson (USA), NSK (Japan) | Ball screws | Planetary roller screws |
| Sensors | Orbbec, Vion Smart, Hesai Tech, RoboSense | Onsemi (USA), Keyence (Japan), Cognex (USA) | Visual sensors, auditory sensors, LiDAR | Six-axis force sensors, inertial sensors |
| Controllers | Unitree Robotics, Xiaomi, Fourier Intelligence | Boston Dynamics (USA), Tesla (USA), 1X Technologies (Norway) | Basic control algorithms | Advanced adaptive controllers |
In intelligent systems, end-to-end multimodal large models are revolutionizing humanoid robot capabilities in perception, decision-making, and execution. Domestically, models like ByteDance’s GR-1 and Peking University’s PixelNav leverage vision-text integration for navigation and task planning, while internationally, Google DeepMind’s RT-2 and UCLA’s MultiPLY excel in multi-sensory interaction. However, foreign leaders often lead in system architecture and algorithm design. For example, the perception-decision-action loop can be represented as: $$ \mathbf{s}_{t+1} = h(\mathbf{s}_t, \mathbf{a}_t) $$ where \(h\) is a learned dynamics model, and \(\mathbf{a}_t\) is the action at time \(t\). This highlights the need for domestic efforts to catch up in core algorithms.
Applications of humanoid robots are still nascent globally, with foreign examples like Tesla’s Optimus in automotive assembly and Amazon’s Digit in logistics, while domestic players like UBTech and Xiaomi are testing scenarios in new energy vehicle inspection. In special applications, such as rescue operations, most remain lab-bound. The table below contrasts intelligent application progress:
| Application Area | Domestic Status | International Status |
|---|---|---|
| Industrial Manufacturing | Phased validation in auto assembly | Deployment in battery plants (e.g., Tesla) |
| Logistics and Transport | Scenario testing in smart logistics | Active use in warehouse tasks (e.g., Amazon) |
| Special Applications | Limited to demonstrations | Experimental use in rescue and security |
| Home Services | Product showcases (e.g., Stardust Intelligence) | Market entry for companionship (e.g., Pepper) |
| Healthcare and Elderly Care | No commercial products yet | Deployment in assistive roles (e.g., Glide robot) |
To advance, we propose focusing on ecosystem-leading enterprises to drive collaborative innovation. By establishing “whole-parts synergy” industrial ecosystems, we can integrate resources for breakthroughs in the humanoid robot’s core components. Initiatives like “unveiling the list and appointing leaders” can cultivate internationally competitive firms that form innovation consortia with academia, addressing key technical challenges. Standardization is vital; we must enhance mutual recognition in the national market and participate in global standard-setting to boost our voice in the humanoid robot domain.
We must seize the “lane-changing overtaking” window by prioritizing the cultivation of operating systems like Open Harmony. These systems, with their natural monopoly tendencies, require early investment to build ecosystems encompassing software, hardware, and services. Supporting national R&D centers and open-source communities can accelerate development. For instance, the value of an intelligent OS can be modeled as: $$ V = \sum_{t=0}^{\infty} \gamma^t R(s_t, a_t) $$ where \(V\) is the cumulative reward, \(\gamma\) is a discount factor, and \(R\) is the reward function for state-action pairs, emphasizing long-term ecosystem benefits.
Guided by industry demand, we should stimulate both supply and demand sides to co-develop application scenarios. Regular matchmaking events and platforms involving users, enterprises, and researchers can identify pain points and guide resource allocation. In mature fields like smart manufacturing, we deepen integration; in emerging areas like elderly care, we launch pilot projects; and in low-tolerance scenarios like disaster response, we tailor solutions for precision and reliability. This approach ensures that humanoid robots meet real-world needs efficiently.
Lastly, we emphasize supporting innovation to overcome technical bottlenecks. Implementing industrial foundation projects can address weaknesses in components like RV reducers and hollow cup motors, while boosting algorithms for decision-making and motion control. Financial and educational policies should foster a collaborative R&D environment, training high-skilled talent to sustain progress. The journey toward mastering humanoid robots is complex, but with concerted efforts, we can lead this transformative industry into a prosperous future.