65. New Year’s Eve Interview with 王兴兴: Behind the Spring Festival Gala and Unitree’s Year

Summary

• 王兴兴 puts the timeline for embodied AI to reach homes and factories at “3 to 5 years in the best case, and absolutely no more than 10 years in the worst case,” and sets his own “dawn” metric: a robot enters an environment that is roughly 80% unfamiliar and completes roughly 80% of the tasks. He says today’s AI models are all pre-trained in advance, so “change even a tiny thing and the success rate collapses”; even the Spring Festival Gala positioning required advance mapping to achieve a high success rate. The day those two 80% thresholds are broadly met, he believes, will be embodied AI’s or robotics’ “ChatGPT moment.”
• The core evidence for Unitree’s motion-control moat is production stability, not demos: before each Spring Festival Gala rehearsal, there was no way to test the robots on site, so the algorithm was tuned at Daxing and could go on stage after an on-site upgrade. Shipping 5,500 units last year meant that “one algorithm had to work on more than 5,000, even 10,000 machines”; early last year, moving the algorithm to another robot made performance “very, very” poor, and only after adaptability improved in the middle to latter part of the year could a 20-plus-robot formation perform together. In-house motors let Unitree modify the motor software directly, which 王兴兴 says “also counts as a moat.”
• Open source is rapidly lifting the industry’s baseline, and 王兴兴’s warning to himself is blunt: “If we had made no product or technology progress over the past year, we would now be a very ordinary, even backward, company.” The host noted that in early 2024, reportedly only 2 companies in the entire industry had robots that could “walk properly” in a Shanghai formation; 王兴兴 agreed that AI advances and open source are quickly raising the floor. Coding Agents will accelerate iteration and shorten the time required to catch up; he agreed when the host said that “most of the code I write is written by AI”—the lead is measured in months.
• On the body-versus-brain debate, he says the embodied AI foundation model matters most, but building the brain carries higher risk, while hardware businesses change more steadily; Unitree’s strategy is to defend its hardware base while continuously filling in the software gaps. His Apple analogy: people think of Apple as a hardware company, but it actually built substantial software and operating systems as well. The embodied foundation model is “the shared global threshold and ceiling”; whoever builds it—“even if it isn’t our company, a Chinese company, or even an American company”—deserves the industry’s thanks.
• He currently favors the world-model route: train a video-generation model together with real-robot trajectories, link data and other information, and it could become a “universal model” that both generates video and makes robots work—a direction that is “very consistent with first principles.” He was challenged last year for questioning the bottleneck in VLA and now says more people “may” agree with him; he also stresses that he changes his mind often: “If I discover tomorrow that VLA is very good, I’ll become a VLA supporter.” The current bottlenecks are poor alignment between video models and physical actions, and multimodal compute costs so high that “a typical pure-play robotics company anywhere in the world may not be able to train it.”
• The ranking investors should remember is simple: technology always comes before orders—“If your technology is good enough, even if you sell not a single robot, you will always be the most powerful robotics and AI company in the world.” The day a genuine embodied AI model breaks through, the market share accumulated in the past will be “fluff in some sense”; he also warns that 2026 could bring overcapacity and vicious competition, urging peers to compete rationally and with restraint rather than wrecking the industry by selling robots below cost.
• Under pressure, his operating philosophy remains unchanged: “Our biggest opponent will always be ourselves”; as long as Unitree keeps iterating, others basically cannot catch up. He also points to the randomness of technology: Google developed Transformer in 2017, but OpenAI only began using it at scale in 2022, and many overlooked technologies may still be buried “under the sand.” That is why he tracks a wide range of developments, including small breakthroughs from individual researchers.

Deep dive

1. “WuBOT” on New Year’s Eve: The Third Spring Festival Gala, Feeding the World’s Martial-Arts Moves to Robots

• The scene opens at 11 p.m. on New Year’s Eve at Unitree’s training facility in Beijing’s Daxing district. By the time the interview questions were asked, it was already 12:30 a.m. on the first day of the Lunar New Year, with a large quantity of medicine on 王兴兴’s desk—he had “had a cold for more than a month.” This was Unitree’s third Spring Festival Gala appearance: robot oxen in the Year of the Ox, yangko dancing in the 2025 Year of the Snake, and 25 robots performing “WuBOT” with young students from Tagou Martial Arts School in the 2026 Year of the Horse. Foreign media called it “the pinnacle of humanoid-robot motion control.”
• The production playbook was straightforward: compile martial-arts moves from around the world, have the robots replicate and learn them, then select dozens for the final performance. The hard part was synchronizing rhythm and visuals: “If the middle is off by 0.1 seconds, we have to fine-tune it, or tear the whole thing down and start over.”
• 王兴兴’s framing of the pressure ran through the entire interview: “Building on what we did relatively well last year, we definitely had to deliver an even better work. The pressure was still very, very high, because we have always known that our biggest opponent is ourselves.”

2. The Show’s Metaphor: For Today’s 10-Year-Olds, the Robot Era Has Already Arrived

• The young performers taunting the robots, the robots performing drunken boxing, and the final hand-in-hand bow were all designed by the production team. 王兴兴 reads the sequence as “human-robot integration, learning from and growing with each other”—a challenge robots must confront if they are to enter daily life.
• His key timing call: today’s 10-year-olds are growing up alongside humanoid robots, and “for them, the robot era has already arrived.” Large-scale deployment into homes and factories will come in “3 to 5 years in the best case, and absolutely no more than 10 years in the worst case.” By the time this generation reaches 20, the technology could be “completely transformed” relative to today.
• He also clarified a trending joke: G1 did not wear pants, while H1—2 meters tall and resembling “the Monkey King”—wore a full outfit. The decision was based on the show’s visual effect and the production team’s judgment. “Sometimes it doesn’t look as good after you put clothes on; sometimes it looks better.” It was not because the robot could not move in them.

3. The Hard Evidence of Global-Leading Motion Control: No Pre-Testing at Rehearsal, One Algorithm Across 5,500 Units

• 王兴兴’s self-assessment was restrained and specific: “In most respects, we are basically the best, although it is possible that individual companies are better than us in certain areas.” His clearest example was the Gala venue, where there was not enough space for testing. Unitree placed dozens of robots both on site and in Daxing; once the algorithm had been validated in Daxing, “a single upgrade” on site was enough to put it on stage, with essentially no major difference in performance.
• The source of that stability is the constraint of mass production. Unitree shipped 5,500 robots last year, so “you are not writing a program for one robot; one algorithm basically has to work on more than 5,000, even 10,000 machines.” The stability bar is “much, much higher than building a demo,” consistent with the industry’s feedback that the robots are effectively “out of the box” ready.
• The requirement runs both ways: hardware must become more reliable and general-purpose, while the algorithm must remain stable “even if the hardware is relatively poor, or has already been replaced.” Early last year, adaptability was still weak: “This robot performed well, but move the algorithm to a slightly different robot and the performance could become very poor.” Adaptability improved substantially by the middle to latter part of the year; otherwise, 1 robot might dance fine while several would fail in a group of 10.

4. Open Source Raises the Floor: A Company That Does Not Improve for a Year Is Ordinary

• The host shared an industry anecdote, with the caveat that it was only “reportedly”: when Shanghai wanted to stage a robot formation in early 2024, only 2 companies in the entire industry had robots that could “walk properly,” and Unitree was one of them. 王兴兴 agreed that AI itself is advancing, then explained the industry’s transformation 6 months later: “Why could so many Chinese robotics companies dance after the middle to latter part of last year? The reason is very simple: open source.” The industry’s baseline “water level is rising.”
• His resulting survival rule was unsparing: “If our company had made no product or technology progress over the past year, we would now be a very ordinary, even backward, company.”
• The host illustrated Coding Agent’s double-edged effect by saying that “most of the code I write is written by AI.” 王兴兴 agreed that the new technology wave will accelerate industry iteration and make catching up faster. The only answer is to “keep improving and keep iterating”; staying number 1 also involves luck, but “once you become complacent, you could be behind within a few months or half a year.”

5. Body or Brain? The Apple Analogy and the Risk Calculation

• Responding to the industry narrative that Unitree is a hardware or “body” company while brain companies command higher valuations, 王兴兴 said: “We are working on both the body and the brain.” Apple is the analogy: it builds its own operating system and software while selling phones. People think of Apple as a hardware company, but it actually builds a great deal of software and systems.
• He did not avoid the ranking: “I have always acknowledged that the most important thing is the embodied AI foundation model. That is unquestionably the most important part of the industry and the world.” The risk calculation is that nobody can guarantee who will build the brain best or fastest; in pure AI, “every few days, one company’s feature is better today and another company’s feature is better tomorrow.” Hardware changes less dramatically: a company that leads this year may still be the leader several years later. Hardware is therefore the base, but the software must also keep improving.
• He does not view the new Locomotion Manipulation trend as new. “We have been building hardware and software for dexterous hands for many years.” Unitree built and open-sourced a video-generation-based embodied model in 2024, making it “one of the earliest companies in the world” to pursue the direction, while also advancing VLA and VLA+RL.

6. The Embodied Foundation Model Is the Industry’s Shared Ceiling: Open Collaboration, and Gratitude to Whoever Gets There

• Unitree’s stance is to explore broadly, invest more in whichever direction works, and collaborate with several domestic embodied-model companies and major internet platforms. “Cooperation does not mean we stop working on our own things.”
• His most expansive statement is worth preserving: “The embodied model is the shared threshold and ceiling for the entire world right now. If anyone in the world builds it—not our company, not a Chinese company, not even an American company—it would be a huge push for the entire industry, and everyone should thank that person.”
• He locates his own sense of achievement in that outcome: the greatest satisfaction of the past few years has been “helping the whole world open-source together and advance the industry together,” which gives him more satisfaction and happiness than how many robots Unitree sold or how high its valuation became. The open-source strategy is pragmatic: “We may keep half of our work closed-source and open-source the other half.”

7. Spring Festival Gala Breakthroughs I: Generalizing Arbitrary Positioning—and the LiDAR Broken by Backflips

• The fundamental difference from peers who simply arrange robots to pose for martial-arts shots was movement: “Start with a formation, run to a new formation, perform martial arts, then rapidly change formation again.” Unitree built a dedicated algorithm for this. The challenge was generalization—not taking a fixed number of steps, but reaching a designated position “regardless of whether it is north, south, east or west,” even after a robot had gone off-axis while performing martial arts. 王兴兴 calls this one of his happiest technical breakthroughs; previously, robots could only “walk over slowly,” which looked bad and had little practical use.
• Localization was the second trap. Unitree used LiDAR, but the robot could not lose its position during backflips and violent movements; if it did, it had to recover immediately. The hardware surprise was that a large number of LiDAR units failed: “These LiDAR units were not designed for violent movements in the first place.” Even after adding shock absorption, service life remained poor: “This LiDAR is still working today, and the next day… it suddenly breaks.” Unitree had to rely on algorithmic correction or other workarounds.
• Launcher-assisted backflips evolved from initially rising “a few tenths of a meter higher than a level-ground flip, still under 1 meter” to roughly 3-plus meters, and at one point the robots “really flew almost as high as the ceiling.” Landing impacts broke the floor and aluminum-alloy components, while aluminum-alloy legs also snapped outright. He treats the Gala as a “major exam”: “It forced us to spend a concentrated period breaking through more cutting-edge technologies—technologies that may be very important but that we might not previously have had time to focus on.”

8. Spring Festival Gala Breakthroughs II: Robot “Hypertension” and the In-House Motor Moat

• On high-altitude landings, the motors absorb energy and feed it back into the battery, causing an abrupt rise in internal current that can damage the battery and other components. The host likened it to blood pressure rising, shock and robot “hypertension.” 王兴兴 said Unitree added special structures to absorb or release part of the energy and prevent sudden high current from damaging the hardware.
• Because Unitree develops its own motors, it can “modify the motor software directly.” The host called that “a moat.” 王兴兴 replied, “Yes, there’s no way around it,” and agreed that “this also counts as a moat.” Players using motors they do not develop themselves “may find it very difficult” to do the same.
• The G1s on stage were the top-spec version: 3 degrees of freedom at the waist, 7 degrees of freedom in each arm including the wrist, and dexterous hands on some robots to improve their grip on staffs. The props included nunchucks, sticks and gourds. The staffs used in the routines with the children were deliberately softened: “Otherwise, if the stick hit or poked a child, that would not be good.”
• The reinforcement-learning training framework was also redesigned around the show. More than 100 movements had to be trained, screened and stitched together quickly. “Many training algorithms in the past could only train one movement at a time, or could not concatenate movements after training,” making overall efficiency very low.

9. “The Hardware Is Good Enough; What’s Missing Is the Model”: The Disabled-Sculptor Argument

• 王兴兴 reiterated his core judgment: dexterous hands are “good enough to use in some sense, though of course not good enough,” with reliability, tactile sensing and other problems still unresolved. But the fundamental issue remains the AI model.
• His argument is the example of people with disabilities: “Some disabled people have no hands, but they can still do their jobs extremely well, and some can even make sculptures. Some people without a single finger can also make excellent sculptures.” If the AI is capable enough, the hardware can be less capable and still get the work done.
• The data methodology mirrors language models: collect movements from many people, clean and filter them, and “not just collect any random movement—you have to collect relatively excellent, representative movements for the model to learn.” The model and the data “complement each other”; having more data does not automatically produce better performance.

10. The Video-Generation Route: Seedance 2.0, the Alignment Problem and Compute That Cannot Be Trained

• Asked whether video generation that obeys physical laws can help train embodied AI, 王兴兴 answered, “Definitely.” Unitree open-sourced a video-generation-based embodied model in 2024; the logic is “very natural and very consistent with first principles”: “If I can generate a video of a robot doing a job, why can’t I use that model to control a real robot?”
• His response to precision concerns is that if the model can even simulate the elasticity of an object during grasping, contact simulation is already reasonably good. Following this route with enough data, “the probability of using data to brute-force an embodied model into existence is still very high.” The biggest current challenge is alignment: video models train on massive datasets while real-robot data remains scarce, so “the video model has already grasped the object, while the physical robot is still a little behind.” A sufficiently large dataset may close that gap directly, with data coming from real-world collection or simulation.
• He also emphasized the limits. Data generated by robots themselves often cannot yet be used directly; people must collect it, or collect it inside simulation. Even if 10,000 robots collected data every day, it is unclear whether the volume would be sufficient, because data quality may not be high. Meanwhile, multimodal models consume such extreme amounts of compute that “a typical pure-play robotics company anywhere in the world may not be able to train them.” Unitree is therefore exploring models that require less data and can complete the loop without massive datasets.

11. VLA vs. World Models: Last Year’s Dissenter Thinks More People May Now Agree

• 王兴兴 was challenged last year after publicly discussing VLA’s bottlenecks. His answer now carries a hint of satisfaction: “Go ask other people now—the number who agree with my view may be higher.” The host also noted a shift at the start of 2026: more discussion of world models, less of VLA.
• His case for world models is that video models already combine language, audio and other inputs into a multimodal model. “As long as you train it together with real-robot trajectories, link data and other information,” the model could become a “universal model” that both generates video and makes robots work. He sees the direction as highly consistent with first principles and expects a major company to pursue it.
• His qualification matters just as much: he currently thinks world models are performing somewhat better, but “there is absolutely no way to say that the VLA route is guaranteed to die.” Unitree is working on VLA itself and will continue optimizing it. The industry is still debating whether VLA is necessary, whether its inference pipeline can be made simpler and more direct, and whether reinforcement learning should be added to VLA.
• His technical method is to stay flexible: “I change my mind often. Today I believe in a direction, we run some tests and experiments, and if we find it does not work, tomorrow we switch to the next direction.” If VLA turns out to be very good tomorrow, he “will become a VLA supporter.” Continuous learning and flexibility matter more.

12. Big-Company Disease, Time Allocation and the “Two 80%s” Dawn

• He uses Transformer to illustrate technological randomness: Google developed the architecture in 2017, but OpenAI only began using it at scale in 2022. Many technologies may still be buried “under the sand,” waiting to be discovered. That is why he reads more new papers and research results every day and keeps his field of attention broad; even a small breakthrough from an individual is worth tracking.
• Unitree is also getting bigger: “There are more people, and management often fails to keep up, so efficiency is actually lower.” Individual intelligence and initiative are not necessarily being fully utilized. There is no optimal solution; the company can only keep improving its management and processes.
• His wish for 2026 is almost radically simple: “Every day, every month, every half-year and every year, there must be continuous product and technology progress.” The breakthrough he most wants to see remains a global embodied model: “Whoever builds it will be good for the entire industry; of course, it would be best if we built it ourselves.” Staying near the top is relatively easy; staying number 1 globally is much harder.
• He sets a quantitative threshold for robotics’ ChatGPT moment: “One day in the future, a robot enters an unfamiliar environment—say, roughly 80% of the environment is unfamiliar—and when you give it a voice command or another instruction, it can complete roughly 80% of the tasks.” That would be close to the “ChatGPT moment” for embodied AI or robotics.
• The industry is nowhere near that threshold. Most AI models today are pre-trained in advance, and a slight change in the setting can send success rates crashing; even Spring Festival Gala positioning requires advance mapping. Reinforcement learning also “has not yet truly delivered its value anywhere in the world,” while Scaling Law and scale have not genuinely arrived. When will the breakthrough come? “Nobody knows right now, but we hope it will.”

13. Warning on Vicious Competition and the Final Ranking: Technology Always Comes Before Orders

• The industry message from the group interview was direct: after resources and capital poured in, “homogeneous competition became extremely serious.” 王兴兴 urged companies to compete rationally and with restraint rather than engage in malicious, wasteful competition. “If everyone starts selling robots below cost, we could wreck the industry.” He also called for patience: this humanoid-robot wave has, “in some sense, only been developing for roughly 3 years,” and locomotion is the prerequisite for robots to do real work.
• Asked about To B and industrial applications, he gave the sharpest ranking of the interview: “The core is still continuous progress in technology and products; everything else is empty. If your technology is good enough, even if you sell not a single robot, you will always be the most powerful robotics and AI company in the world. The day a genuine embodied AI model breaks through, everything you built in the past will have no value compared with it.” When the host summarized it as “technology always comes before orders,” he answered, “Definitely.”
• The host closed by turning the camera on the engineers. She relayed that 蔡明 had said in a CCTV interview that 松延动力’s training facility in Changping had no heating, engineers suffered from chilblains on their hands and slept at an average of 4 a.m. Unitree’s facility was somewhat better, but on the day of the recording, engineers returning from the stage boiled instant noodles, packed up equipment at high speed and rushed home for the New Year.
• Asked about his dream, 王兴兴 gave the same answer as on day 1: “I have loved technology since I was a child. I hope to build excellent technology products and push society forward.” He believes the robotics industry “can truly move human civilization up to a new level.”