Unitree Robotics is becoming one of the most closely watched technology companies in the recent A-share market.

On October 26, Unitree Robotics (688836.SH) continued its sharp rally, rising more than 9% at one point in morning trading and reaching a high of 682 yuan per share. By midday, the stock was trading at 674.01 yuan, giving the company a total market capitalization of 2.726 trillion yuan. Earlier, on October 23, Unitree Robotics had hit the 20% daily limit. Since October 16, it has gained more than 60% over seven trading days.

Behind the market enthusiasm is growing optimism about advances in humanoid robot technology. On October 21, Unitree Robotics unveiled its H2-X humanoid robot, demonstrating complex moves such as platform jumps and forward rolls. The following day, Unitree founder Wang Xingxing said at the World Physical AI Conference that, as large models continue to improve, physical AI is showing signs of making key breakthroughs.

As Unitree’s share price continues to rise, Nomura Securities published its latest global robotics report on October 25, putting forward a seemingly counterintuitive view: as robot motion-control algorithms become easier to train, the factors that truly determine competition in the industry may no longer be motion software, but hardware consistency and the ability to manufacture at scale.

The view is based on a new study of SuperMimic published by the University of California, Berkeley. The technology reduces the computational requirements for training humanoid robots’ whole-body motion controllers by more than 1,000 times. A task that previously required 128 GPUs running continuously for a week can now achieve better benchmark results with just three hours of training on a single H100 GPU.

The research team also deployed the controller on a real Unitree G1 humanoid robot, which successfully performed complex movements including a side flip and a somersault.

This means that robot movement capabilities, which once required enormous computing resources, are becoming increasingly inexpensive to achieve.

Nomura believes, however, that this does not necessarily mean leading robot companies will lose their competitive advantages. On the contrary, as more companies become able to train high-performing motion-control algorithms, another capability may become scarcer: the ability to mass-produce humanoid robots with reliable performance and highly consistent hardware parameters at sufficiently low cost.

According to Nomura’s industry research, around 300 companies in China can already deliver a key motion-control function. Unitree’s one- to two-year lead over the industry in hardware and motion control may be narrowing. However, its accumulated expertise in hardware costs, supply chains, and mass production remains a competitive moat that is difficult to replicate quickly.

For Unitree, whose share price has been rising recently, Nomura’s report offers an investment perspective worth considering: the next phase of competition in the robotics industry may not be about who can make a robot perform the most complex movements, but about who can reliably reproduce those movements across tens of thousands of low-cost products.

One GPU matches a cluster: robot training costs fall 1,000-fold

On October 20, a research team at the University of California, Berkeley, published a paper titled “SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model.” The paper has been accepted to the robotics learning conference CoRL 2026. Its authors include renowned roboticists Pieter Abbeel and Koushil Sreenath.

The research focuses on a central challenge in humanoid robotics: how to teach a robot a wide range of body movements without training costs rising sharply as the number of movements increases.

Training humanoid robot motion control traditionally relies on reinforcement learning. Researchers first build a robot model in a physics simulator, then have the robot repeatedly try different movements. Through extensive trial and error, it gradually learns to walk, run, dance, and perform somersaults.

Although this method is effective, it consumes substantial computing resources. As the number of movements used for training increases, the simulator has to perform more calculations, driving up the cost of training.

For example, training the existing motion-control system SONIC used 128 GPUs for about seven days, consuming roughly 21,000 GPU hours in total.

SuperMimic changes how robots are trained.

Instead of making robots rely entirely on physics simulators for repeated trial and error, researchers first train a model that describes the physical dynamics of the robot itself, then let the motion controller learn within that model.

The model can predict how the robot’s body will move in response to different actions, replacing many conventional physics simulations.

The results show that after three hours of training on a single H100 GPU, SuperMimic outperformed SONIC’s publicly released model in motion tracking. On a benchmark containing 129,514 motions, SuperMimic failed on just 25, compared with 179 failures for SONIC. After a further 9.8 hours of training on two GPUs, SuperMimic had zero failures on the benchmark.

This means the new method can achieve competitive motion-control performance with just a few to a few dozen GPU hours, compared with the tens of thousands of GPU hours required by previous training pipelines.

Nomura says this is among the clearest evidence to date that training humanoid robots for whole-body motion control is becoming inexpensive. More importantly, the technology has moved beyond simulation and onto a real robot.

The research team deployed the trained controller on a Unitree G1 humanoid robot, which successfully performed complex movements such as a side flip and a 360-degree somersault. This shows that motion-control policies trained in simulation using the new method can transfer to real hardware—at least for the range of movements demonstrated in the experiment.

For the robotics industry, this efficiency boost could lower the barrier to developing motion-control technology. Training tasks that previously required companies to have extensive GPU resources and specialized algorithm teams could become feasible for smaller teams as well.

Why is training up to 1,000 times more efficient?

SuperMimic’s key breakthrough is not adding more training data, but changing how robots learn to move.

Nomura points out that humanoid robots use the same body structure to perform different movements, so they do not need to relearn the full physics of their bodies for every individual movement.

The research team capitalized on this by building a robot-body model that can be shared across a wide range of movements. Whether the robot is walking, dancing, or tumbling, the controller can learn within this model how to adjust its joints, reducing reliance on conventional physics simulations. According to the paper, SuperMimic’s efficiency gains come mainly from three factors.

The first is speed. At a typical batch size, the learned world model runs more than 10 times faster than a conventional physics simulator; at the largest tested batch size, it is up to 62 times faster.

The second is the training method. Conventional reinforcement learning relies on extensive random trial and error: the robot repeatedly tries different movements and adjusts its control policy based on feedback.

SuperMimic’s world model is differentiable, allowing it to provide the controller with more direct gradient information and help it determine how to adjust its movements, rather than relying entirely on random exploration. This change alone reduced the number of training steps required by about 64 times.

The third is scalability. When the motion data used for training increased from 1 hour to 264 hours, the world model’s prediction error remained relatively stable. This means that even as the number of movements a robot needs to master grows, the amount of computation required for training does not have to increase proportionally.

Together, the three efficiency improvements reduced GPU-hour requirements by more than 1,000 times with SuperMimic.

However, Nomura also emphasizes that the training conditions for SuperMimic and SONIC were not exactly the same. SuperMimic used about 2.3 times less motion data and a newer GPU model, so the entire difference in computing requirements cannot simply be attributed to algorithmic improvements. The paper’s authors argue that these factors are still not enough to explain such a large gap in computational efficiency.

SuperMimic also does not mean that humanoid robots have achieved general intelligence.

So far, the research has focused mainly on body movements on flat ground, not complex terrain or object manipulation. The world model was trained entirely on simulator data, rather than data from physical interactions in the real world.

The movements in the large-scale benchmark were also part of the training set, so zero failures on the test set does not mean the robot can handle every unfamiliar movement and complex environment.

Nomura describes this as progress at the robot’s “cerebellum” level: a breakthrough in control capabilities for body movement and balance. It is not a breakthrough in the “brain,” which handles complex task understanding, planning, and reasoning, or in the “hands,” which are responsible for precise manipulation.

In other words, SuperMimic greatly reduces the computational cost of teaching a robot to move; it does not solve all the technical problems involved in enabling a robot to carry out practical work autonomously.

Nomura: Unitree’s real advantage may no longer be motion algorithms

More than SuperMimic’s technical breakthrough itself, Nomura is focused on its impact on competition across the robotics industry.

Motion-control algorithms have traditionally been a key competitive moat for humanoid robot makers. Enabling a robot to walk, jump, run, or even perform difficult stunts reliably often requires years of algorithm development, extensive computing resources, and repeated engineering adjustments.

But SuperMimic suggests that this barrier may be coming down. If motion-control training that once required a massive GPU cluster can now be done with a single GPU, more companies will have the opportunity to develop high-performance motion-control systems.

According to Nomura’s industry research, around 300 companies in China can already deliver a key motion-control function. The report says Unitree’s one- to two-year lead in hardware and motion control may be narrowing, while advances such as SuperMimic will further lower the barrier to developing motion algorithms.

Nomura does not, however, believe that Unitree will therefore lose its competitive edge. On the contrary, as motion-control software becomes more readily available, Unitree’s accumulated expertise in hardware manufacturing may become even more important.

The problem is that an algorithm enabling one robot to perform complex movements does not necessarily mean it will run reliably on a thousand robots. Humanoid robots typically have more than 20 joints, each involving actuators, transmission mechanisms, sensors, and precise kinematic calibration. Even two robots built to exactly the same design may have slight differences in actual production.

For example, the joint zero points, dimensional tolerances, and transmission performance may vary between units. When these discrepancies are distributed across more than 20 joints, they can affect the movement of the robot as a whole.

Nomura notes that the robot model used by SuperMimic in simulation had joint calibration errors of only ±0.01 radians. This requirement for hardware parameter consistency highlights the challenges robots face when moving from the lab to mass production. Suppose a company develops an excellent controller that enables a robot to perform a somersault, but hardware differences among mass-produced robots mean that the controller runs reliably on only a handful of units. The company would still have to spend considerable time calibrating and adjusting each robot individually, or even retraining it.

In this situation, even if the algorithm itself becomes much cheaper to train, the costs of deployment and maintenance caused by hardware inconsistency may remain high. Nomura believes that solving this problem requires expertise built up over time in hardware design, supply-chain management, quality control, and mass production. Such capabilities cannot be acquired quickly by buying GPUs or reproducing a research paper. This is an important reason why Nomura believes Unitree retains a competitive advantage over some robotics model companies that build their own hardware.

As the cost of training motion software continues to fall, competition in the robotics industry may shift from algorithmic capability toward hardware precision, product consistency, cost control, and the ability to deliver at scale.

Actuators account for up to 75% of costs: who stands to benefit?

The importance of hardware consistency also means that the distribution of value across the robotics supply chain could change.

Nomura estimates that actuators account for about 50% to 75% of the component cost of a humanoid robot, making them a major part of the total hardware cost. Actuators convert control commands into actual joint movements. Their precision, response speed, stability, and manufacturing consistency directly affect a robot’s ability to perform complex movements.

As motion-control algorithms become easier to obtain, differences in software capabilities between robot makers may narrow, while the manufacturing quality and cost advantages of core hardware such as actuators may become more prominent.

This means that component suppliers with the ability to manufacture high-precision actuators, stable supply chains, and rigorous quality-control systems may be among the key beneficiaries of the robotics industry’s growth.

For makers of complete robots, the key to competition in the future may no longer be whether a single robot can perform a difficult movement, but whether they can deliver consistent performance across large production runs while controlling costs.

Nomura therefore believes that robot companies and key suppliers with strong capabilities in hardware consistency may have an advantage over companies that rely solely on motion-control software.

This view is significant for Unitree. As the cost of developing motion algorithms falls, more competitors may catch up with the lead Unitree has built through its motion-control capabilities. But its accumulated expertise in hardware design, cost control, and mass production may become a more difficult competitive moat to replicate in the next phase.

It is important to note that SuperMimic shows that the computational efficiency of training for specific whole-body motion-control tasks can be greatly improved; it does not mean that robot motion software will immediately become fully standardized. Deploying robots in real-world environments, system integration, safety control, and product reliability will still require ongoing engineering investment.

Nomura’s view is therefore better understood as a projection of how competition may evolve, rather than a claim that motion-control software has lost its commercial value.

Shares up more than 60% in 7 days: what does Unitree need to prove next?

Turning to the capital markets, Unitree’s recent share-price performance shows that investors are reassessing technological progress and commercialization prospects in the humanoid robotics industry.

On October 21, Unitree unveiled its H2-X humanoid robot, demonstrating more complex dynamic movements. On October 22, Wang Xingxing expressed optimism about breakthroughs in physical AI at the World Physical AI Conference.

Unitree shares then hit the 20% daily limit on October 23 and continued rising by more than 9% in morning trading on October 26. By midday, the company’s total market capitalization had reached 272.6 billion yuan.

But advances in robot motion control do not necessarily mean that manufacturers of complete robots will see their profitability rise at the same time. The efficiency gains demonstrated by SuperMimic could allow more companies to enter the motion-control field at lower cost, intensifying competition. Only companies that can turn technological advantages into better products, cost advantages, and the ability to deliver at scale will have a chance to build a sustainable commercial moat.

This distinction is particularly important for Unitree, whose share price has risen rapidly in recent weeks.

On the evening of October 20, Unitree issued an announcement regarding unusual fluctuations in its stock trading. The company said that as of that day, its trailing price-to-earnings ratio had reached 388.25, substantially higher than the average of 39.51 for the general equipment manufacturing industry to which it belongs.

The company also cautioned that the humanoid robot industry remains in the early stages of commercialization, with uncertainty around technological maturity, downstream demand, and the supporting industrial supply chain. If large-scale commercialization falls short of expectations, future revenue growth and earnings realization could be affected.

From an investment perspective, the technological shift represented by SuperMimic raises at least two questions that warrant continued attention.

First, as the cost of training humanoid robot motion-control systems continues to fall, can leading companies maintain their differentiation?

Second, as more companies master complex movements, which manufacturers will be able to turn their technical capabilities into reliable commercial returns through hardware consistency, supply chains, and mass-production experience?

Nomura’s answer leans toward the latter.

As it becomes cheaper to teach robots to walk, jump, and even perform somersaults, what may prove truly difficult to replicate is no longer a particular motion-control algorithm, but a manufacturing system capable of reliably building these capabilities into tens of thousands of robots.

For Unitree, this may be the competitive advantage most worth watching in its next phase—and a key question that will need to be continually tested to justify the market’s high valuation.