Unitree Technology is becoming one of the most closely watched tech companies in the recent A-share market.
On October 26, Unitree Technology (688836.SH) shares continued to surge, rising more than 9% in early trading and touching a high of RMB 682 per share. As of the midday close, the stock was at RMB 674.01, with a total market capitalization of RMB 272.6 billion. Previously, on October 23, Unitree Technology had just recorded a 20% limit-up. Since October 16, it has gained more than 60% over seven trading days.
Behind the market enthusiasm is rising expectations for advances in humanoid robot technology. On October 21, Unitree Technology unveiled the H2-X humanoid robot, demonstrating complex maneuvers such as high-platform jumps and forward rolls. The next day, Unitree founder Wang Xingxing said at the World Physical AI Conference that with the continued improvement of large model capabilities, physical AI has shown signs of achieving key breakthroughs.
Just as Unitree's stock continued to rise, Nomura Securities released its latest global robotics research report on October 25, offering a seemingly counterintuitive judgment: as robot motion control algorithms become easier and easier to train, what truly determines the industry's competitive landscape may no longer be motion software, but hardware consistency and large-scale manufacturing capability.
This judgment comes from the latest SuperMimic research published by UC Berkeley. The technology reduces the training compute required for full-body motion controllers of humanoid robots by more than 1,000 times. Tasks that previously required 128 GPUs running continuously for a week can now be completed with just one H100 GPU and three hours of training, while achieving better benchmark results.
The research team also deployed this controller on a real Unitree G1 humanoid robot, successfully completing complex maneuvers including side rolls and flips.
This means that robot motion capabilities that previously required massive computing investment are becoming increasingly cheap.
However, Nomura believes this does not necessarily mean that the competitive advantages of leading robot companies will disappear. On the contrary, when more companies are able to train excellent motion control algorithms, another capability may become even scarcer: whether they can mass-produce humanoid robots with stable performance and highly consistent hardware parameters at a sufficiently low cost.
According to Nomura's industry research, about 300 companies in China are already able to deliver a certain key motion control function. Unitree's previous one-to-two-year lead over the industry in hardware and motion control may be narrowing. But its accumulation in hardware costs, supply chain, and mass-production experience remains a competitive barrier that is difficult to replicate quickly.
For Unitree, whose stock price has been steadily rising recently, this Nomura research report offers an investment perspective worth noting: the next stage of competition in the robotics industry may no longer be about who can make robots perform the most complex movements, but rather who can reliably replicate these movements across tens of thousands of low-cost products.
One GPU Replaces an Entire Cluster, Robot Training Costs Drop a Thousandfold
On October 20, a research team from UC Berkeley published the paper "SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model." The paper has been accepted by the robot learning conference CoRL 2026, with authors including renowned robotics scholars Pieter Abbeel and Koushil Sreenath.
The research focuses on a core challenge in the humanoid robotics field: how to enable robots to master a large number of different body movements while avoiding training costs that escalate sharply as the number of movements increases.
Traditional humanoid robot motion control training typically relies on reinforcement learning. Researchers first build a robot model in a physics simulator, then have the robot repeatedly attempt different movements, gradually learning to walk, run, dance, and perform flips through extensive trial and error.
While effective, this approach consumes enormous computational resources. As the number of training movements increases, the simulator must perform more computations, driving up training costs.
Taking the existing motion control system SONIC as an example, its training process used 128 GPUs over approximately seven days, consuming a total of about 21,000 GPU hours.
SuperMimic changes the way robots are trained.
Rather than having robots rely entirely on physics simulators for repeated trial and error, researchers first train a model that describes the robot's own physical motion laws, then have the motion controller learn within this model.
This model can predict how the robot's body will move after taking different actions, thereby replacing a large number of traditional physics simulation processes.
Experimental results show that after training for three hours on a single H100 GPU, SuperMimic's motion tracking performance surpassed SONIC's published model. In a benchmark test containing 129,514 movements, SuperMimic failed on only 25 movements, while SONIC's failure count reached 179. After further training with two GPUs for 9.8 hours, SuperMimic's failure count on this test set dropped to zero.
This means that, compared with past training pipelines that required tens of thousands of GPU hours, the new method can achieve competitive motion control performance with only a few to a few dozen GPU hours.
Nomura believes this is one of the clearest pieces of evidence to date that training for full-body motion control in humanoid robots is becoming inexpensive. More importantly, this technology has already moved from simulated environments to real robots.
The research team deployed the trained controller onto the Unitree G1 humanoid robot, successfully achieving complex maneuvers such as side rolls and 360-degree flips. This shows that motion control policies trained in simulated environments through the new method can, at least within the range of motions demonstrated in the experiments, be transferred to run on real hardware.
For the robotics industry, this efficiency improvement could change the development threshold for motion control technology. Training tasks that in the past could only be undertaken by enterprises with substantial GPU resources and professional algorithm teams may in the future be completed by more smaller-scale teams.
Why can training efficiency increase a thousandfold?
The key breakthrough of SuperMimic is not adding training data, but changing the way robots learn motions.
Nomura points out that because humanoid robots use the same body structure when performing different motions, there is no need to relearn the complete laws of physics for every single motion.
The research team leveraged this characteristic to build a robot body model that can be shared across a large number of motions. Whether the robot is walking, dancing, or rolling, the controller can learn within this model how to adjust joints, thereby reducing reliance on traditional physics simulation. According to the paper, SuperMimic's efficiency gains mainly come from three aspects.
The first is running speed. At typical batch sizes, the learned world model runs more than 10 times faster than traditional physics simulators; at the largest tested batch size, the speed improvement reaches up to 62 times.
The second is the training approach. Traditional reinforcement learning relies on extensive random trial and error, with the robot needing to continuously try different motions and then adjust its control policy based on feedback.
The world model used by SuperMimic has differentiable properties, which can provide more direct gradient information to the controller, helping it determine how to adjust motions without having to rely entirely on random exploration. This single change alone reduced the required number of training steps by about 64 times.
The last is scalability. When training motion data increased from 1 hour to 264 hours, the world model's prediction error remained relatively stable. This means that even as robots need to master more and more motions, the training compute does not have to increase proportionally.
The combination of three efficiency improvements enabled SuperMimic to achieve a reduction of over 1,000x in GPU hours.
However, Nomura also emphasized that the training conditions for SuperMimic and SONIC were not entirely identical. The former used approximately 2.3x less motion data and also employed newer GPU models, so the entire computing power gap cannot be simply attributed to algorithmic improvements. The paper's authors believe these factors are still insufficient to explain such a massive difference in computational efficiency.
Furthermore, SuperMimic does not mean that humanoid robots have already achieved general intelligence.
Current research primarily covers body movements on flat ground and does not involve complex terrain or object manipulation. The world model is trained entirely using simulator data rather than real-world physical interaction data.
The motions in large-scale benchmark tests are also part of the training motions, so zero failures on the test set do not mean the robot can already handle all unknown motions and complex environments.
Nomura defines this as progress at the level of the robot's "cerebellum"—that is, a breakthrough in the control capabilities responsible for body movement and balance—rather than the "brain" responsible for complex task understanding, planning, and reasoning, nor the "hands" responsible for fine manipulation.
In other words, what SuperMimic dramatically reduces is the computational cost of teaching robots to move, not the resolution of all technical problems required for robots to autonomously complete real work.
Nomura: Unitree's True Advantage May No Longer Be Motion Algorithms
Compared to SuperMimic's technical breakthrough itself, Nomura is more focused on its impact on the competitive landscape of the robotics industry.
In the past, motion control algorithms were an important competitive moat for humanoid robot manufacturers. Enabling robots to walk stably, jump, run, and even perform highly difficult acrobatic maneuvers typically required long-term algorithm R&D, large-scale computing resources, and repeated engineering debugging.
But SuperMimic shows that this moat may be eroding. If motion control training that previously required massive GPU clusters can now be accomplished with a single GPU, then more and more companies will have the opportunity to obtain high-performance motion control systems.
According to Nomura's industry research, approximately 300 companies in China are already able to deliver a certain key motion control function. The report believes that Unitree's one-to-two-year lead in hardware and motion control may be narrowing, and technological advances such as SuperMimic will further lower the barrier to entry for motion algorithm development.
However, Nomura does not believe that Unitree will lose its competitive advantage as a result. On the contrary, as motion control software gradually becomes more accessible, Unitree's accumulated expertise in hardware manufacturing may become even more important.
The issue is that an algorithm that enables a single robot to perform complex movements does not mean it can run stably across thousands of robots. Humanoid robots typically have more than twenty joints, each involving actuators, transmission structures, sensors, and precise kinematic calibration. Even if two robots use completely identical designs, subtle differences may exist in the actual production process.
For example, different devices may have variations in joint zero points, dimensional tolerances, and transmission performance. When these errors are distributed across more than twenty joints, the motion performance of the entire robot may deviate.
Nomura points out that the robot model used by SuperMimic in simulated environments has a joint calibration deviation of only ±0.01 radians. This requirement for hardware parameter consistency highlights the challenges robots face when moving from the laboratory to large-scale manufacturing. Suppose a company develops an excellent controller that enables robots to perform backflips, but due to hardware differences between mass-produced robots, this controller can only run stably on a small number of devices—the company would still need to invest substantial time in calibrating each unit, debugging, and even retraining.
In this case, even if the training cost of the algorithm itself drops significantly, the deployment and maintenance costs caused by hardware inconsistency may still be very high. Nomura believes that the key to solving this problem lies in long-accumulated experience in hardware design, supply chain management, quality control, and large-scale manufacturing. These capabilities are difficult to acquire in a short period by purchasing GPUs or reproducing a paper. This is also an important reason why Nomura believes Unitree still holds a competitive advantage over some robot model companies that build their own hardware.
As the training cost of motion software continues to decline, competition in the robotics industry may gradually shift from algorithmic capability in the past to hardware precision, product consistency, cost control, and large-scale delivery.
Actuators Account for Up to 75% of Costs—Who Will Benefit?
The importance of hardware consistency also means that value distribution across the robotics industry chain may change.
Nomura estimates that actuators account for approximately 50% to 75% of humanoid robot component costs, making them a critical part of overall hardware costs. Actuators are responsible for converting control commands into actual joint motion, and their precision, response speed, stability, and manufacturing consistency directly affect a robot's ability to perform complex movements.
As motion control algorithms become increasingly accessible, software capability differences among robot manufacturers may narrow, while the manufacturing quality and cost advantages of core hardware such as actuators may become more prominent.
This means that component companies with high-precision actuator manufacturing capabilities, stable supply chains, and rigorous quality control systems may become important beneficiaries of the robotics industry's development.
For complete machine manufacturers, the key to future competition may no longer be whether a single robot can perform a certain highly difficult maneuver, but whether they can maintain consistent performance across large-volume products while controlling costs.
Nomura therefore believes that robotics companies and key suppliers with high-level hardware consistency control capabilities in the future may have an advantage over companies that rely solely on motion control software.
For Unitree, this judgment is of significant importance. As the cost of developing motion algorithms continues to decline, the leading position Unitree established through its motion control capabilities may face challenges from more competitors catching up. However, its accumulated expertise in hardware design, cost control, and large-scale manufacturing may also become competitive barriers that are harder to replicate in the next phase.
It should be noted that SuperMimic has demonstrated that training computational efficiency for specific whole-body motion control tasks can be substantially improved, but this does not mean that robot motion software will immediately become fully standardized. Real-world deployment, system integration, safety control, and product reliability still require sustained engineering investment.
Therefore, Nomura's judgment is closer to an inference about the direction of future competition rather than a belief that motion control software has already lost commercial value.
Stock price surges over 60% in 7 days — what does Unitree need to prove next?
Returning to the capital markets, Unitree's recent stock price performance indicates that investors are reassessing the technological progress and commercialization prospects of the humanoid robotics industry.
On October 21, Unitree unveiled the H2-X humanoid robot, showcasing more complex dynamic motion capabilities. On October 22, Wang Xingxing expressed positive expectations for breakthroughs in physical AI technology at the World Physical AI Conference.
Subsequently, Unitree hit the 20% daily limit on October 23, and continued to rise over 9% in early trading on October 26. As of the midday close, its total market capitalization had reached RMB 272.6 billion.
However, it is worth noting that advances in robot motion control technology do not necessarily mean that complete machine manufacturers' profitability will improve in tandem. The efficiency gains revealed by SuperMimic may allow more companies to enter the motion control field at lower cost, thereby intensifying market competition. Only companies capable of translating technological advantages into product performance, cost advantages, and large-scale delivery capabilities will have the opportunity to further establish sustainable commercial barriers.
For Unitree, whose stock price has risen rapidly recently, this distinction is particularly important.
On the evening of October 20, Unitree issued an announcement on abnormal stock trading fluctuations. The company stated that as of that day, its rolling price-to-earnings ratio had reached 388.25 times, significantly higher than the average level of 39.51 times for the general equipment manufacturing industry to which it belongs.
The company also cautioned that the humanoid robot industry is still in the early stages of commercialization, with uncertainties regarding technological maturity, downstream demand, and supply chain supporting capabilities. If the pace of large-scale commercialization in the industry falls short of expectations, future revenue growth and earnings release may be affected.
Therefore, from an investment perspective, the technological changes represented by SuperMimic raise at least two questions that warrant continuous observation.
First, as the training costs for humanoid robot motion control technology continue to decline, can leading companies maintain their differentiated advantages?
Second, as more and more companies master complex motion capabilities, which manufacturers can rely on hardware consistency, supply chains, and mass production experience to convert technological capabilities into stable commercial returns?
Nomura's answer leans toward the latter.
When it becomes increasingly cheap for robots to learn to walk, jump, and even perform flips, what is truly difficult to replicate may no longer be a particular motion control algorithm, but rather the manufacturing system that reliably installs these capabilities into tens of thousands of robots.
For Unitree, this may be its most noteworthy competitive advantage in the next stage, and will also be the core issue that the market's high valuation needs to continuously validate.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
