Has Japan Lost the "Humanoid Narrative" or the Right to Define the Next Generation of Robots?
Abstract: Japan was once one of the earliest countries globally to deploy humanoid robots, yet it has not become the protagonist in this wave of AI humanoid robot fever. The issue is not "technological backwardness": Japan still possesses profound accumulations in industrial robots, precision reducers, servo control, sensors, and manufacturing systems. The true shortcomings are more concentrated in VLA, robot foundation models, data, software ecosystems, capital, and interdisciplinary talent. With the rise of Physical AI, Japan is attempting to reconnect its "physical advantages" from the industrial era to AI. As humanoid robots move from demos to mass production, can Japan seize the next boom?
From ASIMO to VLA: What Japan Missed Is Not the Humanoid Form, but "AI-Defined Robots"
If we turn the clock back over twenty years, Japan was almost synonymous with humanoid robots globally. Honda's ASIMO brought bipedal robots into the public eye, followed by Toyota launching platforms like the T-HR3, and Kawasaki Heavy Industries long developing Kaleido. Meanwhile, companies like FANUC and Yaskawa Electric established a deep foundation in the industrial robot market. Japan also boasts a complete industrial chain encompassing reducers, servo motors, bearings, sensors, and precision machining. For a long time, "robot powerhouse" has been one of the most distinct labels for Japanese manufacturing.
However, in the current wave of humanoid robot enthusiasm since 2023, the protagonist on center stage has changed. The United States has brought large models, VLA, world models, GPUs, and venture capital into robotics, while China relies on its complete supply chain, manufacturing scenarios, and rapid iteration capabilities to mass-produce humanoid robot integrated units and advance factory validation. In contrast, Japan's presence in humanoid robot startups, financing heat, and AI robot product definition has noticeably declined. It is understood that from 2023 to 2026, a batch of representative achievements in VLA, world models, and robot foundation models were mainly driven by US universities and tech companies. Although Japan has research forces like Preferred Networks, Toyota Research Institute, and Sakana AI, it has not formed a wave of integrated machine startups and model industrialization on the same scale.
This easily creates the impression that Japan "started with a great advantage but ended up trailing behind."
But concluding from this that "Japan can no longer make humanoid robots" is clearly too simplistic. The fact that ASIMO did not evolve into a general-purpose humanoid robot in today's sense does not mean previous technological accumulations were wiped clean. Honda has since continued to advance related robot technologies, extending capabilities like balance control, perception, and whole-body motion to projects such as the Avatar Robot; in 2025, when introducing ASIMO OS, Honda explicitly stated that even though the development phase of the ASIMO robot itself has passed, the company is still pushing forward related robot technologies. Toyota has also not stopped robot research; projects like T-HR3 continue to explore teleoperation, force feedback, and robots assisting humans, while Kawasaki Heavy Industries' Kaleido continues the bipedal and whole-body control route. In other words, Japan has not suddenly lost the ability to manufacture humanoid robots.
What truly sets this wave of the robot industry apart from the ASIMO era is that AI has shifted from an auxiliary technology to the starting point for robot product definition.
Past humanoid robots followed a typical mechatronics route: first perfecting the mechanical structure, motors, reducers, and control systems to solve how to stand, walk, achieve precise motion, and ensure safety control, before gradually adding vision, recognition, and autonomous capabilities. The robot was first a complex and reliable machine, and then gradually became smarter bit by bit.
The logic in the era of large models, however, has begun to reverse. VLA, world models, imitation learning, and reinforcement learning enable companies to first ask "what tasks can AI learn," and then redesign the robot's body around the perception, action, and data required by the model. Robot R&D is shifting from "mechanics defining intelligence" to "intelligence conversely defining mechanics." The importance of models, data, computing power, simulation, and software platforms is rising rapidly, and for the first time, robots are increasingly resembling an AI terminal capable of continuous iteration driven by data.
This is precisely the shift that Japan failed to catch in the first instance during this wave.
Over the past few decades, Japan has built a highly mature industrial robot innovation system: emphasizing precision, reliability, safety, lifespan, and customer ROI, with products requiring rigorous validation before delivery. In contrast, the current embodied intelligence industry is still in a stage of high-speed trial and error, where competition emphasizes model capabilities, data accumulation, software iteration, and capital investment. In the industrial robot system, "minimizing errors" is a core engineering objective; but in the robot learning stage, allowing the machine to continuously trial and error, acquire data, and then improve the success rate through model iteration is precisely part of technological progress. There is a natural tension between these two R&D logics.
Therefore, what Japan has truly missed may not be the "humanoid robot" itself, but the industrial shift where AI redefines robots for the first time, along with the first round of product definition rights it brings.
This leads us to the next question: will this lag persist?
The author believes that as humanoid robots move from demos to factories, commercial spaces, and homes, the competition will ultimately return to precision, motion control, reliability, safety, lifespan, cost, and scaled manufacturing. The capabilities Japan has accumulated over decades in industrial robots, precision machinery, and core components have not become obsolete; Japanese companies are simply waiting for an opportunity.
With Japan's Industrial Robots Being So Strong, Why Can't These Capabilities Be Directly "Translated" to Humanoid Robots?
Looking solely at the industrial foundation, Japan should be one of the most competitive countries for humanoid robots. Data from the International Federation of Robotics (IFR) shows that Japanese companies account for about 38% of global industrial robot production, remaining the world's largest robot producer. In 2024, Japan added approximately 44,500 new industrial robots, with an operating stock of about 450,500 units, and a manufacturing robot density reaching about 446 units per 10,000 employees; in the automotive industry, this figure reached 1,531 units per 10,000 people in 2023.
Data from the Japan Robot Association (JARA) also indicates that this industry is still growing. In 2025, the order value of Japan's robot industry reached 1.0456 trillion yen, a year-on-year increase of 25.7%; production was about 207,000 units, with a production value of 945.3 billion yen, a year-on-year increase of 21%.
This competitiveness stems from a complete robot industrial cluster. FANUC is a global leader in industrial robots, with cumulative shipments exceeding 1 million units in 2023; Yaskawa MOTOMAN reached 500,000 cumulative shipments in 2021, with its advantages extending to servos and motion control; Kawasaki Heavy Industries has long been deeply engaged in automotive welding, painting, and heavy-duty scenarios; Epson excels in small, high-speed, and high-precision robots, and according to Fuji Economy's statistics based on 2024 monetary value, its industrial SCARA robot shipments rank first globally. Mitsubishi Electric's advantage lies in combining robots with PLCs, servos, controllers, and FA systems. Coupled with reducers, motors, bearings, ball screws, and sensors, what Japan has formed is not a single enterprise advantage, but an entire industrial robot ecosystem.
However, being strong in industrial robots does not mean these capabilities can be directly "translated" to humanoid robots.
Traditional industrial robotic arms are usually fixed to the ground, dealing with determined workpieces and trajectories, with core metrics being high rigidity, high precision, high repeatability, and long-term stable operation. Humanoid robots, on the other hand, must carry dozens of joints, motors, reducers, batteries, and computing units on their bodies, while simultaneously solving problems like lightweight design, battery life, dynamic balance, compliant control, and fall impact. If the reducers, servos, and structures used in traditional industrial robots are directly transferred, they will easily face issues of excessive weight, energy consumption, and volume.
The control methods have also changed. Industrial robots mainly handle multi-axis synchronization on a fixed base, whereas humanoid robots belong to a "floating base" system; lifting a leg, reaching out, or carrying heavy objects will all change the body's center of gravity, requiring real-time coordination of dozens of degrees of freedom across the legs, waist, torso, and arms.
"Humanoid robots are all the rage today, seemingly making great progress, but actually making humanoid robots perform relatively large movements is simple. The real difficulty lies in the details—that is, the technology to achieve fine movements in the last inch, millimeter, or centimeter. In this regard, Epson has its own characteristics; we possess excellent control technology," Yasunori Yoshino, head of Epson's robot business, told eefocus. "AI handles higher-level commands and intelligence, but whether the robot can ultimately execute tasks in the physical world still depends on precise sensing and control."
Therefore, Japan's problem is not that its industrial robot advantages have disappeared, but that the migration method has changed. Servos, reducers, precision machining, bearings, and sensors remain important foundations, but joints need to become lighter and more compliant, controllers need to interface with ROS 2, GPUs, and AI, and traditional teaching and rule-based programming must further integrate with VLA, reinforcement learning, and Sim2Real.
Decades of accumulation in Japan's industrial robots have left a very good "body." But whether this body can adapt to the AI era and complete the transition from precision machinery and motion control advantages depends on Physical AI capabilities.
Reducers, Ball Screws, Bearings, and Sensors: How Much Advantage Remains in Japan's Upstream Industrial Chain?
If Japan's presence in the current wave of humanoid robot integrated machine enthusiasm has declined, the situation in the upstream industrial chain is much more complex.
In the field of robotics, the advantages Japan truly cannot bypass include a complete set of precision component systems: Harmonic Drive Systems' harmonic reducers, Nabtesco's precision reducers, THK's linear motion components, NSK and NTN's bearings, Nidec's motors, as well as visual, inertial, and force-sensing devices covered by Sony, TDK, Murata, and MinebeaMitsumi. These capabilities remain an important foundation for Japan's entry into humanoid robots.
Reducers are the most typical example. Industrial robot joints need to convert the high-speed rotation of motors into low-speed, high-torque, high-precision motion, making harmonic reducers and RV reducers long-standing strengths of Japanese companies. In the era of humanoid robots, precision and lifespan remain important, but since robots must carry dozens of joints on their bodies, reducers must also be lighter and smaller, while withstanding impacts from walking, jumping, and even falling. The barriers established in the past through "high precision, high rigidity, and long lifespan" still exist, but they need to be redesigned around lightweighting, impact resistance, and low friction. It should be noted that existing data is not yet sufficient to prove that these companies have dominated the humanoid robot supply chain; actual orders and revenue contributions still need to be confirmed in conjunction with corporate financial reports.
Ball screws and bearings are facing similar changes. The ball screw, linear guide, and precision bearing technologies accumulated over the long term by THK, NSK, and NTN can enter mechanisms such as linear actuators, knee joints, and elbow joints, where high thrust density, low friction, and reliability remain important. Information shows that THK has also begun to launch related modules aimed at lightweight and compact needs.
What has truly changed is the value chain structure.
In the industrial robot era, manufacturers could separately procure motors, reducers, encoders, bearings, and drivers, and then complete system integration; however, humanoid robots are increasingly tending to compress these components into highly integrated joint modules. Integrated machine manufacturers are no longer just concerned with how high the precision of a single reducer is, but rather the torque density, weight, energy consumption, heat dissipation, cost, and maintenance convenience of the entire actuator.
This poses a new challenge for Japanese suppliers. In the past, pushing a single component to the extreme could secure a high-value segment; in the future, if competition shifts to "integrated actuators," single-item performance advantages may be diluted by system integration capabilities. Especially in China's humanoid robot industrial chain, the matching degree between traditional standard servo products and highly customized joint modules is declining, while the proportion of local suppliers and integrated machine manufacturers developing their own drivers is rising.
The same applies to the sensor field. Sony has deep accumulations in CMOS image sensors, TDK and Murata have technical foundations in MEMS and inertial sensing, and MinebeaMitsumi covers micro-motors and force-sensing capabilities. However, visual sensors themselves are already a fully competitive mature market; entering the robot BOM does not equate to mastering the definition rights of the perception system. Higher technical barriers may appear in six-axis force sensing, tactile arrays, and the deep integration of sensors with joint control and AI models.
Therefore, Japan's status in the upstream of humanoid robots can be summarized as: individual hardware remains very strong, but competition is moving from "components" to "modules and systems."
Advantages of Japan's Robot Industrial Chain, Source: Compiled by eefocus Research Institute
This is also the most critical turning point for Japan's robot industry next: it still possesses a very strong "body," but whether it can upgrade its component advantages into system capabilities will directly determine its position in the humanoid robot era.
Japan's True Shortcomings: From AI and Data to Interdisciplinary Engineers
If a robot is divided into a "brain" and a "body," Japan's contrast today is quite obvious. On the "body" side, Japan still possesses profound accumulations in industrial robots, precision reducers, motors, servos, sensors, and manufacturing processes; but the other set of capabilities that have driven the rapid development of humanoid robots in recent years—foundation models, VLA, world models, robot data, GPU computing power, and software ecosystems—Japan's overall presence is weaker than that of the US and China.
This does not mean Japan lacks AI research. Toyota Research Institute, Preferred Networks, Sakana AI, as well as universities and research institutions, continue to invest in machine learning and robotics, and influential achievements such as Diffusion Policy have emerged. The more obvious gap is that these technologies have not yet formed a dense "model-data-computing power-robot body-capital-application" cycle like that in China and the US.
One important reason stems from the software system. Traditional Japanese industrial robots have long been built around deterministic control; PLCs, dedicated controllers, teaching pendants, and trajectory planning pursue real-time, reliable, and verifiable performance. However, the technology stack required for embodied intelligence is increasingly approaching ROS 2, Python, PyTorch, CUDA, reinforcement learning, and large-scale simulation. The resulting disconnect in software interfaces, control architectures, and development paradigms has become a major obstacle to migrating traditional robot capabilities to Physical AI.
The data issue is even more typical. Japan has a large number of robots operating long-term in automotive, electronics, and precision manufacturing production lines, which seemingly should provide rich physical world data. However, what traditional industrial robots save are often motion trajectories, equipment status, and I/O signals, whereas VLA and robot foundation models require synchronized vision, language, joint status, force sensing, tactile sensing, human teaching, and records of success and failure. Coupled with the fact that factory data is often owned by customers, and formats and protocols vary among different manufacturers and equipment, Japan's robot operating experience has not naturally transformed into standardized multimodal data that can train large models.
This forms a highly noteworthy "paradox of an industrial powerhouse": Japan may possess extremely rich robot experience, but it has not timely transformed this experience into data assets for the AI era.
Institutions like NEDO have recently begun building robot data infrastructure and foundation model projects, indicating that Japan has recognized this issue. Traditional robot companies are also beginning to adjust their directions; manufacturers like FANUC, Yaskawa, and Kawasaki are increasing investments in AI, simulation, and open software platforms, attempting to connect past motion control systems to Physical AI.
But ultimately, it is engineers who will truly execute this transformation.
Japan does not lack talent in mechanics, servos, motion control, and embedded systems; what is truly scarce is a cross-disciplinary capability: understanding both robot dynamics, joint friction, structural resonance, and motor control, while being able to use PyTorch, CUDA, reinforcement learning, VLA, and Sim2Real to train robots. The author reviewed recruitment information from relevant Japanese manufacturers and found that traditional companies like FANUC and Yaskawa still lean more towards motion control, mechanical design, and embedded systems, whereas Chinese and US embodied intelligence companies are more intensively recruiting talent for RL, VLA, foundation models, CUDA, and data engineering.
Therefore, what Japan truly needs to make up for in the next stage is neither simply adding a few AI positions nor just adding a few algorithm courses for traditional engineers, but rather restructuring the entire robot R&D process so that models, data, simulation, control, joints, and the main body can iterate synchronously.
Japan already possesses strong "Physical" capabilities. Whether it can re-enter the center of robot competition largely depends on whether it can cultivate a batch of "new species" engineers who stand between AI and the physical world—capable of tuning both models and robots—and thereby complete the transition from an industrial robot powerhouse to a Physical AI powerhouse.
Has Japan Really Missed Out? The Real Race May Have Just Begun
Returning to the initial question: Has Japan really missed out on humanoid robots?
If the "boom" is defined as the integrated machine releases, financing, VLA models, and humanoid robot demos since 2023, Japan is indeed a beat behind. Whether it is the capital and model frenzy sparked by US companies like Figure AI and Physical Intelligence, or the rapid iteration of Chinese humanoid robot companies in integrated machines, supply chains, and application scenarios, Japan has not been the most active player in this stage.
But the truly difficult part of the humanoid robot industry may have just begun.
The competition in the past few years is more like a "0 to 1" race: can the robot run, complete grasping, and understand natural language instructions; can the model show a certain degree of generalization ability on new tasks. In this stage, AI models, GPU computing power, financing capabilities, and rapid trial and error are especially important, which happens to be the area where US and Chinese companies are more active. We can summarize the industrial characteristics from 2023 to 2026 as robot demo competition.
But when robots enter actual application scenarios for deployment, a beautiful demo is far from enough.
Robots must answer a series of very traditional, and even somewhat "boring," engineering questions: How many hours can it run a day? Does joint precision decrease after continuous work for several months? Can it still work after a fall? What is the lifespan of the motors, reducers, and drivers? Is the failure rate acceptable? How long does maintenance take? Can a robot ultimately be cheaper than manual labor or traditional automation equipment?
These questions are precisely what Japan's robot industry has been solving for decades.
The product logic formed by traditional industrial robot manufacturers like FANUC, Yaskawa, Kawasaki, and Epson often prioritizes reliability, precision, safety, maintenance costs, and customer ROI. This approach may seem conservative in the early stages of rapid trial and error in embodied intelligence, but it may become important again after humanoid robots enter factories. The capabilities accumulated by Japan's upstream companies in reducers, bearings, motors, sensors, and precision machining are also more likely to demonstrate their value in this stage.
Therefore, Japan's opportunity in the next stage is not simply to repeat the success of the industrial robot era.
Humanoid robots require lighter joints, more compliant control, and more open software architectures, as well as VLA, robot foundation models, and large-scale data training. Japan's past advantages must complete at least three transformations: transforming the precision engineering of industrial robots into lightweight and compliant humanoid robot capabilities; transforming the experience accumulated at manufacturing sites into data that can train AI; and then connecting the "Physical" advantages composed of reducers, servos, and sensors with VLA, world models, and open software ecosystems.
If these three transformations cannot be achieved, Japan may ultimately remain just a high-end component supplier for the next generation of robots, handing over the rights to integrated machines, models, and industrial definition to Chinese and US companies.
Finally, the author believes that Japan has indeed missed the first wave of the humanoid robot boom driven by AI models, integrated machine demos, and venture capital, but this boom is not over yet.
Over twenty years ago, Japan proved that robots can walk like humans; in this current wave, Chinese and US companies are proving that robots can understand tasks and learn movements. Next, the question is whether robots can work as reliably as human workers while reducing costs to a level that truly holds commercial value.
This race has just begun.
