90. Why Are Humanoid Robots So Fixated on Track Meets? | An Interview with 熊友军 of Tiangong
Summary
熊友军 defines humanoid robot track meets as an extreme testing ground: the competitions are not a substitute for industrial deployment, but a way to find the limits of the body, power system, motion control and reliability against a common yardstick. Tiangong Omni improved from 9.44 seconds in rehearsal to 8.64 seconds, as the density of competition kept pushing both software and hardware toward their limits; the real commercial threshold is far tougher—factories demand 7×24 operation and, in some cases, a 99% success rate. You have to be willing to sit on a cold bench for 10 years.
The viral “faceless run” was not a preprogrammed performance, but a solution the robot found on its own while chasing maximum speed under shoulder-joint and power constraints. The team initially disliked the posture as “not that pretty” and even considered retraining it, but kept the fastest version: the robot gave up arm swing and used its waist to balance the lower body. 熊友军 sees this as an instance of embodied-intelligence emergence: “It really was a result of self-learning and self-evolution.”
The most important annual change at the track meet was not speed but autonomy: last year, Tiangong was the only one of more than 200 robots to compete fully autonomously; this year, track-and-field robots had largely left the remote controls behind except in some obstacle events. 熊友军 sees this only as “a demonstration of progress in the brain,” because recognizing lane markings, taking bends and adjusting running form remain relatively shallow capabilities; long-horizon planning, error reflection and handling sudden environmental changes have yet to be fully tested.
Embodied intelligence may ultimately require data on the scale of tens of millions of hours, but 熊友军 refuses to treat “throwing more at it” as the inevitable end state. Today’s general-purpose motion controllers use only hundreds of hours of human data, while world-model training runs to roughly tens or hundreds of thousands of hours—still several orders of magnitude below 10 million hours. He admits, “We genuinely still don’t know” where the scaling-law ceiling lies, and says that relying purely on data accumulation “may not necessarily be the right approach”; autonomous exploration, reinforcement learning and few-shot learning could still change the required scale.
The real threshold for training a robot’s “brain” is a combination of compute, data and learning paradigm—not the launch of another model name. 熊友军 agrees that 1,000-GPU effective compute is a meaningful benchmark; with Baidu and multiple cloud platforms, the innovation center collects nearly 200,000 hours of data a year, has surpassed 20 million downloads, and is shifting its focus from repetitive pick-and-place tasks toward long-horizon tasks, logical chains, force and tactile sensing, failures and counterexamples. Data-driven methods are already deployable, but VLA, world models, multimodality and unified models have yet to converge.
Near-term commercialization will still begin in structured industrial settings, but those settings are better understood as a bridge to household robots than as the final destination. In parcel sorting, the fastest humans handle more than 2,000 items an hour, FANUC was cited at more than 1,500, and domestic robots at 1,000 to 1,400; the gap reflects a combination of body design, algorithms, models, data and scenario customization, with no obvious generational lead yet visible at WRC demonstrations. 熊友军 remains convinced that humanoids will ultimately become “friends, partners, even family,” while the host keeps her reservations: production facilities may instead be redesigned around “people + robots + AI,” with the humanoid form not necessarily the only end state.
The core investment question is not who raises the most money during the boom, but who has durable technical compounding, a clear set of advantages and the willingness to make it through the trough. 熊友军 says the industry is “essentially not competing yet” because the blue ocean is large enough, but capital can buy isolated wins, not years of foundational accumulation; the innovation center is building public infrastructure around the Tiangong body, Huisi Kaiwu’s “brain” and a global open-source ecosystem. “Choose a direction … and then make time your friend” is the answer he has carried for more than 20 years.
Deep dive
1. The “faceless run” was the optimal solution forced by constraints
卫诗婕 asked why Tiangong Omni bobs its head, barely swings its arms and yet runs so fast. 熊友军 explained that during training, the team deliberately constrained the shoulder joints’ rotation frequency, range, speed, angle and force output, leaving reinforcement learning to find its own way to maintain balance.
Humans typically use arm swing to counterbalance lower-body movement, but swinging the arms also consumes power. To optimize the overall objective of completing the 400m and 1,500m runs at the fastest possible speed, Omni shifted the compensation to its waist, producing the posture that users dubbed the “faceless run.” “It really was a result of self-learning and self-evolution.”
The team initially disliked the posture and thought it “wasn’t that pretty,” even considering training a better-looking movement. But it was the fastest of all the options, so performance won out. 熊友军 sees it as a localized emergence in embodied intelligence: engineers did not write the movement directly; the strategy emerged from the interaction of objectives, rewards and physical constraints.
2. From 9.44 seconds to 8.64 seconds, competition keeps lifting the ceiling
Tiangong’s pre-race rehearsal time was about 9.44 seconds. After seeing the competitors’ performances, the team continued iterating on its algorithms and testing the hardware ceiling, eventually reaching 8.64 seconds. 熊友军’s view is that competition forces every team to “keep breaking through, keep trying,” drawing out capacity that would not normally be tested.
The improvement did not come from a single algorithm. The team prepared multiple strategies in simulation, then used real-world testing to filter them. 熊友军 emphasizes that “AI is essentially an experimental science”: it can borrow from human training methods, but it also runs different approaches and lets the results determine what survives.
He said the championship was “basically expected,” but the repeated running videos reaching the top of the trending lists were entirely unexpected. He sees the attention as public recognition of the team’s years of accumulated motion-control work, not as evidence that the underlying technical path suddenly changed.
3. The value of the meet is extreme testing, not repetitive spectacle
Responding to the criticism that the industry has already moved toward industrial work and should not still be competing in running and jumping, 熊友军 gave a direct definition: the meet is an “extreme testing environment” for extending the boundaries of the body, reliability, stability, mobility and energy system.
The work demonstrations at the WRC booth represent the industry’s move from laboratory spectacle to the “main battlefield of the national economy,” but they remain some distance from real production. Factories need 7×24 operation, and production lines may require a 99% success rate. The meet provides a common yardstick and training ground for companies to test their limits together.
His causal chain is straightforward: competitions first expose weaknesses in the body, joints, energy system and control algorithms; those foundational capabilities then support entry into factories, malls and homes. Even the best “brain” cannot execute without a reliable body—much like the relationship between the brain and the hands.
4. The real annual leap was robots putting down the remote controls
熊友军 recalls that more than 200 robots competed at the previous meet, but Tiangong was the only team to compete fully autonomously; most other teams had people running alongside with remote controls. This year, apart from occasional remote intervention in some obstacle events, the track-and-field robots were largely autonomous.
Fully autonomous running requires sensing lane markings, identifying bends, deciding when to lean into them, and adjusting running form and speed strategy as conditions change. 熊友军 sees this as a concentrated demonstration of greater investment in the embodied “brain,” not simply a faster execution of the old program.
But he also draws a firm boundary around the achievement. The brain capabilities displayed at the meet remain limited: long-horizon task planning, reflection after execution errors and responses to sudden changes in complex environments “have not yet been demonstrated.” Future events should gradually add these tests.
5. The general-purpose motion controller starts with human data, but rewards and physics determine the ceiling
The team used motion-capture equipment to record human running form at different speeds, arm swing, balance strategies, and changes in the ankles and knees across different phases. It then fed hundreds of hours of data into the model for basic imitation. 熊友军 said NVIDIA Sonic used more than 700 hours of human data, broadly in line with the industry’s current scale.
Imitation is only the starting point. The “most important” elements remain the reward function, penalty function, optimization target and constraints. Maximum joint torque, speed, angle and output capacity define the physical boundary, and the control policy must solve for the global optimum within it.
The controller itself is not large; it is a vertical-domain model. It remains far from the performance humans expect. The current controller mainly covers whole-body movement, especially the lower body, while more complex manipulation and dexterous-hand control still lack sufficient data.
After winning its first marathon, the team open-sourced the motion-control code and training pipeline. 熊友军 said the same paradigm can also be used to train complex movements such as Thomas flares and backflips.
6. A scaling law may exist, but nobody knows the answer for embodied intelligence
Asked whether a scaling law would require 100 million or 1 billion hours of data, 熊友军 refused to offer a seemingly precise number. He would only say that the 10-million-hour scale is certainly necessary, while today’s motion controllers use only hundreds of hours and world models roughly tens to hundreds of thousands of hours.
He acknowledged: “We genuinely still don’t know.” The uncertainty concerns not only the data volume but also the eventual paradigm for embodied world models and robot brains. VLA, 4D models, multimodality plus VLA, world models and unified models have all yet to converge.
卫诗婕 took the question one step further: even after a paradigm converges, the data required by the new paradigm may not be the same as the data everyone is collecting today. 熊友军 agreed, meaning current data assets cannot be valued linearly by hours alone; task structure, modalities and learning mechanisms will also determine their value.
7. “Throwing more at it” is a necessary solution for now, not necessarily the end state
熊友军 describes the massive-data route as “a solution we currently have no choice but to use.” Large language models can draw on the internet’s existing stock of data, while embodied data is expensive, scarce and tied to interaction with the physical world. Simply copying the scaling path of language models may be neither economical nor consistent with how humans learn.
His counter-question is: “Humans can generalize from just a few actions, or a small amount of data … so why do machines necessarily need that much data?” If autonomous exploration or a new few-shot paradigm breaks through, zero-shot or few-shot generalization could reduce the data requirement.
This is not a rejection of the current path. 熊友军 believes VLA already works in relatively structured factories with controlled environments and clearly defined tasks. As complexity rises to human-machine interaction, unexpected changes and long-horizon tasks, multimodal models and world models will need to handle more simulation and reasoning.
8. The data race is moving from UMI and Ego toward long-horizon logic and multimodal feedback
熊友军 does not treat UMI, Ego or full-body teleoperation as inherently right or wrong. Once the industry discovers that a particular data type trains models effectively, it forms a consensus that lasts roughly six months; when supply is no longer scarce and limitations become visible, collection shifts to the next mode.
Full-body teleoperation requires real robots, motion-capture suits or controllers, and extensive scene setup, making it expensive and slow. Ego only requires workers to wear first-person cameras, allowing it to cover rich scenarios cheaply and quickly. But first-person video still cannot fully capture force, touch, execution logic or the reasons for failure.
The innovation center’s next focus is long-horizon tasks and their logical chains: “Why do this first, this second and this third?” Data should also include speech, vision, force, touch, error sequences and counterexamples, so models learn not only to reproduce movements but also to reason and correct mistakes.
9. “Why pick up the trash?” exposes the missing value layer in data-driven learning
熊友军 uses picking up trash to illustrate the gap in current training. A model can be told what counts as trash and where the bin is, then execute the grasp-and-drop motion; but it does not know why leaving the trash there is harmful, nor does it naturally develop an initiative to keep the environment clean.
Human actions may be driven by values and a worldview, not merely by sensing an instruction and reproducing a movement. 熊友军 notes that scholars internationally are already exploring whether robots should have some form of value system. He does not assert an answer, but believes these questions will push embodied learning into philosophical territory.
His approach is to walk on “two legs”: industrialize VLA, multimodal models and world models where they can be deployed, while continuing to “look up at the sky” for learning mechanisms closer to autonomous cognition. Staying grounded and exploring new paradigms are not contradictory in his view.
10. Sutton-style autonomous exploration offers another low-data path
熊友军 said Richard Sutton visited the innovation center to exchange ideas. His reinforcement-learning approach does not continuously feed robots large volumes of human demonstrations; it sets the environment, rules and direction of iteration, allowing robots to “explore, receive feedback, correct, learn and improve.”
He compares the process to a PDCA spiral: after a failure, the robot corrects itself based on feedback and enters the next round of exploration. The key resources shift partly away from massive demonstration datasets toward high-quality environments, reward mechanisms and the ability to sustain trial and error.
熊友军 retains his uncertainty: “It may take some more time to explore exactly which one can ultimately work.” But he believes this kind of autonomous iteration could surpass pure imitation, which is also why the strange running posture learned by Omni itself matters.
11. Moving from predictable mathematical models to learning systems was a paradigm break
Before 2010, the team relied primarily on mathematical models for motion control. Engineers planned every step, solved the equations and issued commands one by one. The movements were predictable and diagnosable, but they turned the robot into “a puppet.” A slight environmental change could make the system “completely stop working.”
Even a simple task could require thousands or even tens of thousands of lines of code. 熊友军 remembers the industry asking whether robotics had already reached its technical limit. Deep learning, large models, vision and natural-language technology later changed the entire solution space.
He compares Honda missing this transition to “Nokia missing the smartphone.” ASIMO’s era pushed mechanics, manufacturing and reliability to the extreme, but remained built on traditional mathematical models, limited vision and limited compute, and failed to capture the shift in AI paradigms.
12. P2 sparked the ideal, but also showed the distance between a technical manifesto and a product path
While pursuing his doctorate, 熊友军 studied welding robots and spent his days dealing with weld uniformity, appearance and adaptation to variation. Around 2002, he saw Honda’s P2 greet visitors and take them around an assembly line. The gap between the two sides was “not even close,” and for the first time he felt directly that “this is the future.”
熊友军 says Honda should have begun investing in humanoid robotics in 1986, using the automotive industry’s advantages in precision components, manufacturing, reliability and stability to become an object of aspiration for an entire generation of researchers. But the robots then carried massive industrial computers, while natural-language capabilities, vision and compute were all inadequate.
卫诗婕 notes that ASIMO was more like a technical manifesto and never entered real production at scale. 熊友军 does not deny its value: many of today’s humanoid visions still do not go beyond the picture Honda drew, but market demand does not mean the technology is mature. NAO and Pepper later reinforced the same point.
13. UBTECH’s early moat came from building its own technology when no supply chain existed
熊友军 joined UBTECH because the team wanted robots to enter the home, closely matching his vision of “human-machine integration.” He describes it as “the feeling of finding someone who understood me.” In the early days, the office had only a few people and a few desks; founders would stay up all night after employees left, discussing the future and then pitching the “dream” to the team and investors.
The first Walker prototype started with two legs because adding the upper body would sharply increase control complexity. The team initially planned to source joints from Germany and South Korea, but had to develop them in-house after the products proved unreliable. ASIMO and Atlas were highly closed systems at the time, leaving no shortcut through buying mature technology.
In 2012 and 2013, the mobile-phone supply chain was built around orders in the thousands and even millions, while humanoid robots were produced 10 or 20 units at a time, with demanding technology and tight cost constraints. The team had to wait for gaps in machine-shop schedules and wine and dine factory managers and machine operators before anyone would take a “small order” seriously.
14. The high points were not sudden viral hits, but the accumulation of 20 years of industry memory
When Unitree drew public attention with its Spring Festival Gala handkerchief routine and martial-arts performance, 熊友军 saw it as “a golden age for the industry” rather than asking why the spotlight was not on his own team. He notes that UBTECH had already put 541 robots on the Gala stage in 2016: 1 interacted with the hosts, while 540 formed three dance formations.
Recalling UBTECH’s difficult years, he thinks first of the Spring Festival Gala, the listing and other high points, saying he has always “enjoyed the process.” From its registration in March 2012 through the ups and downs leading to the IPO, the lasting lesson was summed up in two words: persistence.
In the early years, capital was chasing the internet, mobile internet and real estate, making robotics financing difficult. His undergraduate studies were close to automotive engineering, and he could easily have shifted into new-energy vehicles, which later attracted huge amounts of capital. But his answer was always, “I just liked it.” “When you like someone, you love everything associated with them”; failures become tests and opportunities to grow.
15. Four team values explain why even a 90 is not good enough
熊友军 set four values for the team: integrity, self-drive, collaboration and excellence. Integrity means putting character before work; self-drive means being willing to work toward a long-term goal without supervision; collaboration follows from robotics’ systems complexity across mechanics, control, algorithms, data and applications.
“Excellence” does not mean scoring 90 or 95 on an individual component. His exact standard is: “If you don’t deliver 120%, you haven’t met the bar.” The robotics value chain is too long; if every link performs at 90%, errors compound until the final result is “a small deviation becoming a thousand-mile mistake.”
Integrity is the hardest quality to quantify, so he looks more at how people view their dreams, troughs and long-term commitments. During the previous humanoid boom, many companies moved into faster-money businesses once the industry cooled. Those who enter at the peak and leave in the trough are unlikely to become the kind of fellow travelers he wants.
16. Teleoperation is not backward technology; pretending it is autonomy is crossing the line
卫诗婕 mentioned that one team was disqualified for claiming autonomy while actually using teleoperation. 熊友军 says that violated both the competition rules and the team’s values, and must be policed by the judges. But he also stresses that teleoperation itself has a legitimate role in high-risk settings.
If a red light appears at a switchgear cabinet or substation, a wrong operation could have severe consequences, so a human should decide whether to flip the switch. Once authorized, the robot can autonomously navigate obstacles, open the door, walk to the equipment and complete the action. This is no longer toy-car-style directional control; it is “the human making the high-level decision while the robot handles the execution.”
He distinguishes micro-action, behavior-level and task-level teleoperation. The first is close to a puppet; the behavior level can issue a local instruction such as “pick up the cup”; the task level combines human judgment with autonomous machine execution. When intelligence is still insufficient, this hybrid approach may be safer than pursuing formal full autonomy.
17. New players can buy isolated wins, but not long-term compounding
On automakers and smartphone companies entering robotics, 熊友军’s first reaction is that it signals an industry boom. Commercial companies “come for the profits,” meaning they believe this is a RMB10T-scale blue ocean: invest RMB100M today and potentially pursue returns of RMB1B, RMB2B or even RMB10B.
The host cited Honor’s April marathon victory and the “red-versus-white battle” on the track, asking whether traditional robotics companies could be overtaken by hardware makers with deeper pockets. 熊友军 recognizes Honor’s technology and investment, but says the only conclusion is that “real work produces results”; it does not follow that humanoid-robot moats have disappeared.
His rebuttal targets the capital market’s linear thinking. Investing RMB500M or RMB1B can indeed produce a short-term result at a particular point, but robotics is not a field that can be completed with a one-time investment. “You have to be willing to sit on a cold bench for 10 years.” Foundational technology, reliability and systems integration require long accumulation; one successful phase does not guarantee success every time.
18. In a blue ocean, there is no true head-to-head competition—only different endowments being realized
熊友军 believes that although there are many embodied-intelligence companies today, “everyone is essentially not competing,” because the unmet market is large enough. Solving a real problem in one vertical can create an independent space, which is also why capital continues to invest.
When evaluating a company, he first asks “why it exists,” then whether its subsequent resource allocation truly follows that mission. He next looks at the team’s endowments: those with industrial and manufacturing expertise can go deep into factories, while those familiar with healthcare and elder care can work on health or exoskeletons. Similar visions matter less than matching the strategy to the capabilities.
His view of followers is clear: companies without an independent position that merely copy others during a hot cycle may ultimately be “drowned in the wave of embodied intelligence.” Persistence does not mean keeping the same product unchanged; it means continuing to invest time, energy and resources in a direction the team genuinely believes in and is good at.
19. Whether the humanoid is the end state remains the interview’s most important disagreement
卫诗婕 proposed an industrial-revolution analogy: facilities were once designed around human labor, then rebuilt around machine power once steam matured. In the same way, future facilities may be redesigned around “people + robots + AI,” meaning today’s humanoid form may not be the final one.
熊友军 agrees that the environment will change, but holds to his foundational view: the future is a society of human-machine integration, where robots become friends, partners and even family. To support that relationship, robots need forms, movements and communication styles that humans understand naturally, and they should be able to use tools and spaces designed for people without requiring wholesale environmental reconstruction.
He calls the humanoid a potential “third-generation human-machine interaction hub.” It would no longer require unlocking a phone and opening an app, nor rely on keyboards, mice and GUIs; it would infer intent through language, gestures and even eye contact. In the tests, research and comparisons his team has conducted, once cost and technical complexity are temporarily set aside, people choose the humanoid over the robot dog.
卫诗婕 said explicitly, “I’m personally not in that camp.” 熊友军, meanwhile, described a “three-body” vision: the physical 熊友军, a robotic 熊友军 and a digital-human 熊友军 existing together, with knowledge and experience potentially transferred through brain-computer interfaces and perhaps creating a form of “immortality.” The disagreement was not forced into a resolution.
20. Industrial deployment is the realistic bridge imposed by technical constraints
Walker initially targeted the home but later prioritized factories. 熊友军 says this reflects both current technical limitations and the need for commercial returns: homes involve tougher safety, standards and open-environment challenges, while structured industrial settings are more controllable and easier to monetize.
Industrial deployment can also generate data in reverse. Once robots enter real frontline operations, they can collect large volumes of data and accelerate iteration across the Walker line. Going industrial first and home later is not abandoning the long-term vision; it is building the commercial returns, reliability and data foundation needed to reach it.
熊友军 has not yet seen an obvious generational gap among the manufacturers at WRC. He cites parcel sorting: the fastest humans handle more than 2,000 items an hour, FANUC was cited at more than 1,500, and some domestic systems at 1,000 to 1,400. Performance depends on the combined capabilities of hardware, algorithms, brain models, data and scenario customization.
Soft packages, hard boxes, plastic bags, different colors, sizes and shapes require large volumes of real-world training. VLA can already generalize to some extent within this limited range of variation, but that is still far from an open household environment with arbitrary objects, arbitrary tasks and human-machine interaction.
21. The national team’s job is to build infrastructure, not just win for one company
The host noted that the Beijing Humanoid Robot Innovation Center was established in November 2023. After moving from UBTECH’s CTO role to become the center’s CEO, 熊友军 says his role shifted from the commercial company’s “small self” to the industry and nation’s “larger self”: not just pursuing commercialization for one company, but mapping the industry’s pain points, hard problems and bottlenecks.
The innovation center positions itself as an infrastructure builder. Tiangong provides a general-purpose body platform, while Huisi Kaiwu covers the small and large brains—the interview earlier also called it “Huisi Kaiwu” under another name. The center continues to open-source hardware designs, the marathon training pipeline, millions of high-quality data records and model architectures, reducing redundant foundational work across the industry.
Early shareholders including UBTECH, Xiaomi, Yizhuang and Jingcheng Machinery share more than an expectation of financial returns; they also recognize the platform’s technical and industrial value. 熊友军 says the open-access principle is “equal for everyone”: developers in the US, Europe, Japan and South Korea can download the resources as well, so he does not see the platform as an exclusive bloc.
Compared with UBTECH’s emphasis on applied technology, the innovation center must invest in deeper common bottlenecks, tools and ecosystems. Large numbers of universities, research institutes and secondary developers use its platform to compete, and 熊友军 sees that developer density as evidence of platform value rather than looking only at medals won by its own teams.
22. The hard threshold for training the brain is 1,000-GPU compute, data quality and sustained investment
卫诗婕 relayed an industry estimate that no more than 5 independent startups without government or major-tech backing truly possess the effective compute of 1,000 GPUs needed to train a brain. 熊友军 did not confirm the number of companies, but agreed with the direction of the benchmark: the compute required for early-stage training is “extremely, extremely high.”
The innovation center initially built its own compute center, but it quickly fell short as data and models grew. After Baidu became an investor, it supplied substantial cloud resources, and the team also uses other cloud services. Beyond training, large-scale deployment will create even greater inference demand, though inference costs can be spread across a growing customer base.
Annual data collection is already close to 200,000 hours, which 熊友军 says is a very large volume by domestic open-data standards; cumulative downloads have exceeded 20 million. High quality does not mean stacking repetitive grasping tasks into more hours. It means covering multiple tasks, bodies and scenarios, long-horizon logic, tactile feedback, failure sequences and counterexamples, with the team training and validating the data.
23. Unified models aim to repair the “last-mile” losses in embodied intelligence
卫诗婕 relayed 王兴兴’s description of the bottleneck: language models input and output within a fixed digital-vector space, so information loss is relatively small; every perception and action by a robot introduces error, and tiny tactile deviations are especially difficult to correct. The longer the task, the faster the success rate tends to collapse. 熊友军 explicitly agrees.
The innovation center’s answer is to map language, vision, touch and other modalities into a unified representation space, then use one model for reasoning and decision-making. 熊友军 says the team has launched the RoboBrain Unify unified large model, ranked first on the authoritative World Arena leaderboard, and also delivered strong results on voice and vision benchmarks.
The approach aims to combine VLA’s strengths in manipulation and execution with the world model’s strengths in reasoning and prediction. The cost is a parameter count far larger than that of a single model; current robot-side chips struggle to support it, so most of the workload still runs in the cloud.
熊友军 expects the model to be reduced through distillation, pruning and architectural changes once it matures, then gradually deployed on the edge. He defines this as “a direction worth exploring,” not as an industry consensus that has already formed.
24. No single company can monopolize the true full stack
Facing the large number of companies that call themselves full-stack, 熊友军 says the term “full stack” needs quotation marks. The robotics technology tree includes joints, key components, motion control, MPC, WBC, multiple brain paradigms, operating systems, toolchains and different types of data. No company can go deepest at every layer.
The innovation center’s own full stack refers only to coverage of several key common technologies around the Tiangong body platform and Huisi Kaiwu’s small- and large-brain platforms. Its teams ranked first in gold medals in competitive events and second in scenario-event golds, reflecting the national platform’s need to scan industry weaknesses—not proof that specialists have no value.
熊友军 instead defends specialization: “Every trade has its master.” A startup that reaches the top in one critical direction deserves applause. What the industry really needs is a large number of companies each solving the technologies they understand best, then combining them through open source and the supply chain—not every company copying a fake full-stack model.
25. China and the US have different advantage structures; China needs to build the original brain
熊友军 reduces the global embodied-intelligence competition to China and the US. The two countries are at least on the same starting line, but their advantages differ: China is stronger in supply chains, manufacturing iteration and the breadth of application scenarios, while the US leads in original AI, brain models, world models, spatial intelligence and compute chips.
China has done extensive work on planetary and harmonic reducers and ball screws for humanoid joints, with stability and reliability continuing to improve. Some well-known overseas robotics companies also source key hardware from China. Electronic skin, pressure sensors and batteries have improved particularly quickly over the past 2 years.
Batteries remain a weak point, but volume, capacity and discharge rate have improved markedly. 熊友军 even predicts that “perhaps at the next marathon, you’ll see a battery run the whole course,” illustrating how supply-chain optimization can directly change the product boundary.
Domestic compute already has replacement options, but they require reworking the underlying architecture and development stack and may not yet be the most economical choice. As long as overseas chips remain available, companies may not voluntarily absorb the switching cost. 熊友军 acknowledges that China is still more of a follower in original areas such as world models and spatial intelligence, and must close the gap through long-term investment.
26. Open source turns public investment into compounding developers, feedback and talent
The innovation center chose open source for more than publicity. The first return is user development: people downloading its code and data today may become customers or ecosystem partners tomorrow. The second is faster iteration through cases, modifications, suggestions and bug reports obtained under open-source licenses.
熊友军 says the center’s dataset is probably the most downloaded in China and may even rank first globally, though he has no confirmation of the global ranking. He cares more about the idea that “many hands make light work”: an open platform can bring developers together, reduce redundant investment in common technologies and train embodied-intelligence talent in China.
Compared with the closed-source route common in the US, he believes China’s open culture can create a different diffusion advantage. Open source does not mean abandoning commercialization; it means enlarging the public base while companies continue to compete in applications, products and services.
27. Safety is not an add-on feature, but the entry ticket for commercialization
The faster and stronger a robot becomes, the more serious the consequences of a fall, collision or structural failure. 熊友军 breaks safety into several layers: the body must not harm people or the environment; data and information must be protected; personal privacy must not leak; and the system must be protected from hijacking.
He expects some regions and scenarios to impose restrictions on robots in the same way they do on drones. Before robots can be sold at scale, they must also meet mandatory national safety regulations. The innovation center is participating in standards work under the guidance of the Ministry of Industry and Information Technology; 熊友军 says multiple standards will be issued “by the end of this year,” and products that fail to meet them will not be allowed to sell.
Asked whether a hacked robot could be prevented from becoming a killer, 熊友军 says encryption and protection already exist for networks, communications and system software, so the problem is not technically unsolvable. The real gap is unified laws, regulations and mandatory enforcement; many startups still do not invest enough in this area.
28. A robot is most alive not when it wins, but when someone needs it
熊友军 first imagines a desktop robot. While he is focused on work, it suddenly stands up and says, “Boss 熊, a visitor has arrived. You need to go meet with them.” The natural voice, intonation and movement made him feel for the first time that the robot was like a secretary—and like a real person.
The second scene takes place in a COVID isolation ward, where a robot delivers daily necessities to people who are helpless. The third is in a nursing home, where a robot leads elderly residents through exercises and their novelty-driven questions keep coming. 熊友军 says these scenes have the warmth of “help arriving in the snow,” making technology produce genuine social value.
卫诗婕 felt a different kind of emotional value from the meet: clumsy robots boxing, running and falling made friends gather around the television and laugh together for the first time in a long while; when the robots broke records, they inspired awe toward “silicon-based life.” Clumsiness and breakthrough together create the pleasure of watching.
熊友军 watches more like a father watching a child learn to walk, with “eyes full of expectation and love.” Omni and Archo often fall during testing; an arm may land crooked, and the chest casing may deform. He personally proposed putting foam protective clothing on key areas. Others see the punchline; he sees a child that falls, needs protection and is still growing.