161: 唐文斌 of 原力灵机: China’s Earliest AI Entrepreneur and His New Embodied-AI Journey
Summary
- 唐文斌 sees the primary bottleneck in embodied intelligence today as “IQ,” not the robot arm, camera, or other hardware. End-to-end control may achieve only 50%-60% performance, while desktop-task success rates on RoboChallenge may be as low as 50%; 原力灵机’s generalist model is still only in the 30%-plus range. The company’s technical priority for 2026 is to materially improve multi-task/generalist generalization, rather than keep piling onto single-task specialists. Of the “intelligent, useful, and trustworthy robot,” the latter 2 are currently blocked first by the first.
- His bar for commercialization is far stricter than “deliver one robot”: a real-world setting, 10-plus hours of operation per day, 2 months of continuous running, and a meaningful fleet size. Pressed on what “meaningful scale” means, he casually cited 100 units; by that standard, he believes there may be “almost no” mature examples in the industry. 原力灵机’s 2026 target is “one setting, 1,000 robots,” using deployment data—not collection data—to start the flywheel. The point is not to scatter across a dozen POCs, but to close the loop in 2 or 3 settings.
- Logistics is the first beachhead not because robots are already 100% capable, but because the system can absorb their mistakes. When a robot fails, the dispatch system can route the task to a human station or hand control to a remote operator; the customer is buying a continuous workflow combining robots, people, other automation, and dispatch. The setting must also offer error tolerance, time tolerance, moderate generalization, and long operating hours before total cost can be calculated.
- 原力灵机 is betting on “embodied-native” training: robot data should enter the joint training mix with internet multimodal data, including video, and autonomous-driving data from day one of VLM pretraining. 唐文斌 compares bolting an Action module onto an existing VLM to sending a middle-school graduate straight into a sports academy: the underlying ceiling is low, and the model may even forget its “Chinese and math.” Action, grounding, VQA, and other capabilities therefore need to be co-trained during the VLA stage as well. DM0 and 阶跃 are jointly completing native VLM training, after which 原力灵机 continues with VLA and post-training.
- On hardware, 原力灵机 does not treat a fully standardized body—including a standard humanoid—as the default answer for every setting; its rule is that “the model unlocks the setting, and the setting defines the hardware.” Standard bodies can be either underkill or overkill, while the true ceiling is first determined by whether “a human brain plus teleoperation” can complete the task. A high-center-of-gravity humanoid or wheeled dual-arm robot may top out at 0.8-1 meter per second, versus 4 meters per second for a low-center-of-gravity AGV; some workstations do not require the robot to move at all. The company’s answer is rapid assembly from standard modules, not a fresh custom design for every project.
- 唐文斌 sees the durable moat not as a temporary lead in models, but as a data flywheel powered by usage. Megvii’s face-authentication business generated more data with every additional authentication, which in turn improved financial-grade accuracy; by his account, it still holds roughly 70%-80% market share. Ordinary SDKs lack that loop, leaving their advantage as merely a time gap. The same applies to embodied intelligence: the most valuable data is ultimately not expensive collection or polished demos, but failures, interventions, and task feedback from robots operating at scale.
- The core lesson from his 15 years in To B is that value must be measurable, products must be standardized, and competition’s tendency to induce overpromising must be treated as a risk. The 2019 Uniqlo project went through 5 rounds of bidding and used more than 1,000 robots, but initially achieved only 40% of its target efficiency; Megvii secured a 6-month extension and rewrote all the code from scratch before delivering. 唐文斌’s conclusion is that “fake ROI” on a PowerPoint does not generate repeat business, while the market’s “if you say you can’t do it, they’ll just think you’re weak” dynamic systematically hides failed projects.
- He expects the industry to first accommodate a large number of vertical companies, then consolidate around a single-digit number of global platform players as general-purpose models absorb smaller models, with most of the eventual winners potentially coming from China. The final qualification will depend on the interaction of model IQ, closed-loop deployments, and data flywheels—not fundraising or current rankings. Risk management means “stockpiling enough provisions” while avoiding shotgun expansion. Megvii’s retrospective formula was that “A plus B plus C is basically equal to max(A, B, C),” which is why 原力灵机 rejects showpiece robots and high-cycle-time, low-tolerance auto plants, concentrating resources on general-purpose models and a small number of scalable settings.
Deep dive
1. Programming First Delivered the Satisfaction of Solving Problems
When 唐文斌 was in eighth grade in Xinchang, Zhejiang, the school had just started organizing informatics competitions and selected the 2 students with the best math scores from each class. Once he and 2 or 3 friends got the key to the computer room, many classmates preferred playing games, but they thought “programming was actually more fun than playing games.”
Math and physics problems generally ask you to find an answer. Programming asks you to design “a method for finding the answer”; the path need not be unique, and optimization problems may not even have an absolute optimum, only successively better solutions. That openness later extended directly into AI, dispatch, and robot optimization.
After a program was submitted, an immediate wrong answer or accepted result created powerful positive feedback. 唐文斌 calls it “the engineer’s thrill of solving problems”: conceive the algorithm, implement it elegantly, pass every test, then extend the capability to websites, tools, and even game cheats.
2. Topcoder Compressed Thinking, Competition, and Instant Feedback into 5 Minutes
唐文斌 continued competing in Topcoder, ACM, and other contests at university, not to secure a recommendation for graduate school, but because “it was fun.” His favorite Topcoder event was the Challenge: after the regular contest, contestants had only 5 minutes to read code from others in the same room and use counterexamples to reduce an opponent’s score to zero while earning 50 points themselves.
Each submission could be Challenged only once, so the event tested not just algorithms but also speed and observation. He remembers one contest in which all 20 people in the room submitted the same problem; the other 19 missed the same small trick, and he broke all 19 programs, earning 950 points from Challenges alone.
The experience shaped his instinct for evaluation: a good system should not merely produce an answer, but expose its boundaries quickly under pressure. Codeforces is now used as a benchmark for code models; in his view, it is still quantifying the same kind of ability.
3. The Core of a Good Benchmark Is Discrimination, Not Difficulty
From 2007 to 2013, 唐文斌 and another IOI gold medalist, 胡伟栋, led student work including contest problem-setting. 唐文斌 continued participating even after spending 2 years on a startup. He says writing a problem may take only 1 or 2 hours, but writing a good one can take a week.
The point of setting problems is to genuinely separate the abilities of roughly 200 contestants: the problem must be original, test a clear concept, and not simply ratchet up the difficulty. “A good problem is not defined by how hard or easy it is. The core requirement is discrimination.”
That experience later transferred to embodied-intelligence evaluation. 唐文斌’s principle is: “You have to be a good player yourself to be a good judge,” because without hands-on R&D and operational intuition, it is difficult to design a benchmark that pulls the technology forward.
4. The Competition Network Became a Talent Chain Spanning More Than a Decade
唐文斌 met 楼天成, 蒋凡, 王小川, and others through the competition circuit, and encountered many more students while coaching and lecturing. 范浩强 joined the national training team when moving from ninth to tenth grade and competed in IOI from tenth to eleventh grade; 唐文斌 recalls that he “should have” placed second. A rule had also been introduced that students who competed internationally and won gold would not continue competing, though 唐文斌 clarified that 范浩强 had already secured an early admission offer in ninth grade.
When 范浩强 asked, “What am I supposed to do now?”, 唐文斌 invited him to join newly founded Megvii. The company first assigned the high-school student to test deep learning because the new technology had no established playbook; someone fast and smart enough to experiment may have mattered more.
唐文斌 had known 范浩强 since the latter was in seventh grade and 赵尔静 since roughly eighth grade, and they worked together for years afterward. Asked whether Megvii had at one point employed more than a dozen IOI gold medalists, he corrected the premise: “We had even more NOI gold medalists.”
5. 印奇’s Organizational Ability and a Shared Interest in Vision Led to a Startup Pact
唐文斌 and 印奇 were not close from the moment they entered Tsinghua. They later ended up in a class of 27 men and no women, and both interned at Microsoft Research Asia. 印奇 worked with 孙健 on face research, while 唐文斌 moved unexpectedly from data mining and social-network research into image search.
A class-wide student-festival production changed 唐文斌’s view of 印奇. 印奇 served as director and organized all 27 classmates around a complex script that was “both warm and funny.” 唐文斌 saw not just technical ability, but the ability to organize a large group of people.
Before graduating in 2009, the 2 agreed to start a company together one day. At the time, it was not yet fashionable to call everything AI; they talked about machine learning, computer vision, and the simple conviction that “all robots will need eyes in the future.”
唐文斌 became 唐杰’s first graduate student after 唐杰 obtained the right to recruit students. During his adviser’s academic leave, he joined a Microsoft Research Asia group that had just shifted from text mining to image search. With the attitude of “I’m here anyway,” he took on the new direction, eventually completing image-search work that was used in Bing.
6. Kinect, iPhone 4, and the Independent-Developer Wave Triggered Megvii
Around 2010, Kinect’s motion-controlled interaction was booming, while the iPhone 4 introduced a front-facing camera to phones for the first time. 唐文斌 and 印奇 wondered whether a Kinect-like experience could be brought to mobile devices now that phones could see faces and hands.
The success of independent developers such as 墨迹天气 convinced them that a small team might also make money from an app. Each invested in a MacBook, then brought in 杨沐 after realizing they lacked development capacity. The 3 built the motion-control game Crows Coming, or 《乌鸦来了》.
The game did not make the money they had expected, but it briefly reached the top 3 of the App Store’s free-app rankings and won industry awards, attracting a wave of investors. The founders had only a vague plan to “start a company in the future,” so they decided that “learning on the battlefield is also a pretty good way to learn” and formally incorporated after taking investment from Lenovo Star and Lenovo Capital.
Megvii later built a 3D roller-skating game in which users jumped with their phones and dodged missiles with their faces; the team also imagined a motion-controlled version of Angry Birds using the phone as a slingshot. They had plenty of technical ideas, but gradually admitted they did not genuinely enjoy designing casual games, monetization hooks, or progression systems.
7. After Abandoning Games, Face++ Became an Open Experiment in Finding a Market
唐文斌 says the team played games occasionally, but preferred larger titles such as StarCraft; they found no satisfaction in the commercial design of casual games. At the end of 2012, Megvii stopped all game development and confirmed that its goal was not to become a game company, but to apply computer vision across more settings.
Face technology was still immature, and the team did not know where the best beachhead lay, so it started Face++ and opened the API to developers for joint exploration. The company supplied face landmarks, face slimming, and eye enlargement SDKs for 美图秀秀 and 美图手机, and also experimented with visual apparel search and clothing recommendations.
The experience exposed a constraint in the founders’ DNA: they lacked a deep feel for both mass To C aesthetics and To B settings, often “looking for nails with a hammer.” 唐文斌 says bluntly that the things they liked “were not mainstream tastes,” and that starting from technology alone made it difficult to become a naturally user-driven company.
8. Alipay Account Opening Turned Deep Learning into a Scaled Business for the First Time
Megvii’s first genuinely scaled business was financial-grade face recognition for Alipay’s online account-opening needs. Customers had to verify that a remote applicant was the actual account holder, and after comparing market offerings, Alipay judged Megvii’s performance to be the best.
唐文斌 recalls that the industry’s LFW performance was around 99.1% at the time; he also notes that Megvii was among the first to move performance directly from the 90%-91% range into the 98%-99% range. The inflection point was the rise of deep learning alongside AlexNet, although the team was initially “half-convinced, half-skeptical.”
The company asked 范浩强 to run the first experiments. 范浩强 and 赵尔静 formed the internally dubbed “强静 combination” and materially improved the full chain from face-landmark detection through final recognition. 唐文斌’s takeaway was not to chase every new concept, but to remain open to new technology and use experiments to reach a conclusion quickly.
9. Only a Data Flywheel Can Turn a Technology Time Gap into a Moat
程曼祺 noted that many technical advantages from the previous AI startup cycle never became commercial results. 唐文斌’s correction was: “It still comes down to the setting.” A pure technology gap is often just a time advantage that others can close within days; the real barrier is the positive loop in which usage generates data and data improves the product.
Identity authentication has that flywheel. After serving Alipay, Megvii expanded into other internet-finance settings, Didi, Uber, and elsewhere; the more authentication events it processed, the more people and edge cases it saw, bringing the system closer to financial-grade accuracy. By 唐文斌’s account, the business still holds roughly 70%-80% market share.
Beauty SDKs are different. Customers do not demand extreme accuracy, nor do they continuously return enough data to create a meaningful lead, so the moat ultimately remains a time gap. Large customers can of course build in-house, but when authentication is only a small part of the business, an external supplier can usually amortize R&D across multiple customers more economically.
10. AI To B Must First Answer Value, Cost Share, and Standardization
唐文斌’s first principle is that customer value must be measurable: exactly how much cost was reduced or efficiency increased. A product cannot stop at a visually attractive PowerPoint ROI. “Fake ROI” may help win the first order, but cannot support repeat purchases or scale.
His second principle is to solve a problem that is widespread but does not account for too large a share of any one customer’s cost base; if the share is too high, the customer is incentivized to build a substitute. 程曼祺 challenged him with Nvidia and CATL: isn’t high share plus irreplaceability even better? 唐文斌 acknowledged that this is powerful, but said automakers building batteries and large companies developing their own chips also show that substitution pressure never disappears.
The third is to avoid the “project trap.” B-end demand is inherently diverse, so a supplier must specify which functions are standardized, which are configurable, and which truly require customization. If every customer gets a new build, revenue may grow, but the organization and delivery model cannot scale.
11. Below-Zero Temperatures, 40 Kilometers, and 200% Attrition Launched the Logistics-Robotics Business
Around 2016, 唐文斌 and 印奇 visited a Tmall Supermarket warehouse in Tianjin. The building was large and unheated, with temperatures around -10°C; pickers pulled carts for roughly 40 kilometers a day, equivalent to running a marathon in the cold.
A manager asked them to guess the annual attrition rate. 唐文斌 guessed 60%; the manager replied, “That shows a lack of imagination.” The real figure was 200%. The problem was not merely the wage bill, but the difficulty of recruiting and retaining workers and the resulting management complexity—issues that were even more acute in Japan, South Korea, Europe, and North America.
The team had dreamed of robots from day one, and the MegV, or Mega Vision, name used by Megvii at the time also carried the idea of first teaching robots to understand the world. The warehouse turned that long-term ambition into a concrete need: give robots “hands and legs” to produce output in the physical world.
12. Standard Warehouse Objects and Dispatch Made Robots Scalable for the First Time
Unlike an open environment, a warehouse constrains objects at least to pallets, totes, or individual items, creating a degree of standardization that makes it a suitable starting point for robots. The team’s long-term vision remained to move beyond warehouses, campuses, and fixed boxes, but general-purpose robots were not technically realistic at the time.
What initially determined system efficiency was not how intelligent any single vehicle was, but slotting, order priority, traffic management, and multi-vehicle coordination. 唐文斌 viewed the warehouse as a large optimization problem, directly matching the team’s background in competitions, operations research, and reinforcement learning.
To move away from pure project work, the business broke the warehouse into receiving, pallet storage, tote storage, shipping, and other modules. Each used a standard product and solution, with configuration handling the combination; integrators could also assemble complete systems from these “building blocks.”
The separated logistics business, 原力聚合机器人, used Atomics as its brand. Revenue was roughly split evenly between projects and products; pallet and tote products, sales, and channels were all maturing. 唐文斌 says scale has reached “several hundred million yuan,” with many additional orders not yet recognized as revenue, while the overseas share is rising quickly.
13. The Uniqlo Project Fully Exposed the Cost of Overpromising
In 2019, the bidding for Uniqlo’s automated warehouse in Shanghai brought in nearly every relevant domestic supplier and went through 5 rounds of screening “like a talent show.” The project used more than 1,000 robots and was extremely complex. Megvii acknowledged that it did not have the most experience, but argued that ultimate efficiency depended on software, algorithms, and dispatch.
To B competition forced suppliers to keep raising their promises: when a competitor said it could do something, the customer asked whether Megvii could do it too, and the team accepted many requirements it itself viewed as carrying extreme technical risk. Once the system was complete, actual efficiency was only 40% of a 100% target. “We almost blew it.”
Megvii explained to the customer that it had identified the problems, requested a 6-month extension, and rewrote the entire codebase from scratch. The project was eventually delivered, and the customer instead recognized the team’s persistence. 唐文斌 later traveled to Japan to meet Uniqlo founder 柳井正, and the company won additional work afterward.
14. Competitive Procurement Systematically Creates Overpromising and Invisible Failures
唐文斌 still has no easy answer. If a supplier is not yet a recognized authority, honestly saying “we can’t do it” is often interpreted by the customer as weakness rather than professional judgment; the competitor that promises first wins the order first.
Feedback cycles often run on a yearly timeline. If a supplier fails 6 months later, the customer may already have lost confidence. Neither side wants failed projects made public, so the market sees successful publicity but not the aggregate cost of overpromising.
Once trust is established, the relationship changes: the supplier can flag risks, both sides can validate them through a small-scale experiment, and move from “no” toward “how to yes.” 原力灵机 hopes the technical trust built with customers through its existing logistics business will reduce the pressure to exaggerate capabilities simply to win an order.
唐文斌 describes many current embodied projects as POCs, PR exercises, or even “staged photo shoots.” Outside entertainment and research, there is no clear answer to how many robots are actually running continuously. Vendors collectively raise customer expectations even when the technology may not yet be ready.
15. Real Deployment Depends on the Full Workflow, Not on Assuming a Single Robot Never Fails
“Useful” first means solving the customer’s complete task, and second means making the economics work. A robot occasionally completing one action does not mean the customer’s process is closed-loop; if a failure becomes an exception requiring someone to be called in ad hoc, the system remains a POC.
唐文斌 does not assume models will reach 100% anytime soon. A more practical product designs the failure paths in advance: route the task to a human station, hand over to a remote operator, or let another device continue the work, so the workflow stays continuous even when the model fails.
Closed-loop operation and ROI are separate questions. The fallback mechanism first ensures the task is not interrupted; then the total cost of robots, human labor, remote operators, dispatch, and equipment is aggregated to confirm that the customer still saves money or increases output.
16. 原力灵机 Is a New Company Restarting After a Complex Restructuring
The Megvii team began experimenting with large-model-driven robot control and VLA in 2024; 原力灵机 formally began operations in spring 2025. The intervening months were spent mainly on restructuring: some shareholders exited, a new fund took over their shares, and existing shareholders retained partial interests in the new company.
范浩强, 蒋竞天, and other core members had no shortage of outside opportunities, but still chose to start another company together. 唐文斌 attributes the decision to years of accumulated trust in one another’s abilities, a sense of brotherhood, and the customers, settings, and engineering base left by the logistics-robotics business—all of which raised the odds of success.
原力灵机 has roughly 100 employees, with about 40% coming from former colleagues and 60% new hires. The company set incentives as a fresh startup and reserved an option pool of roughly 30%, rather than placing new arrivals into an old structure that had already hardened.
17. High Talent Density Accelerates Iteration but Cannot Complete Commercialization on Its Own
唐文斌 believes “strong people like having stronger people around”: even when they disagree, talented colleagues recognize one another’s independent thinking and ability to solve hard problems. That chemistry reduces procedural coordination and speeds up R&D iteration.
Talent density does not automatically become revenue. Commercial deployment requires not only solving the hardest algorithms, but also understanding customers, handling dirty work, and filling in every operational detail. Those jobs require different types of people working together, not simply another batch of competition gold medalists.
A saying once circulated at the Megvii Research Institute: “That’s not fundamental.” Researchers quickly lost interest in problems that were not sufficiently difficult. 唐文斌 now makes the opposite point: “When you truly become a commercial product, everything that affects the customer is fundamental.”
Financial returns were not romanticized. He admits that he and 印奇 “feel guilty” because after years of leading the team through startups, they had not enabled everyone to make money. A team’s higher aspirations for technology and social value do not mean economic returns are unimportant.
18. Repeatedly Underestimating Technological Change Led 唐文斌 to Back Embodied Intelligence Earlier
Megvii initially doubted that deep learning would be much better than SVM and similar methods. After Transformer emerged, 张祥宇 published “Make CNN Great Again” in an effort to show that CNNs could achieve comparable results; the team eventually had to acknowledge that Transformer was evolving faster.
Autonomous driving went through the same early skepticism and is now on the streets. 唐文斌’s postmortem is that “every technological shift in history has been underestimated in terms of how quickly it would change things.” Even as the psychological threshold rises, technology often moves faster than expected.
Embodied models currently have very low success rates, but they are on a “very steep improvement gradient.” The influx of capital, talent, and attention will further accelerate the experiment loop. To him, robots represent a kind of ultimate vision for AI producing labor in the physical world; the demand itself does not need to be created.
19. “Intelligent, Useful, and Trustworthy” Ultimately All Point to Model Capability
原力灵机 defines its mission as “building intelligent, useful, and trustworthy robots.” “Intelligent” identifies the technical variable behind the new venture; “useful” requires solving the full task and making the economics work; “trustworthy” includes safety and reliability.
唐文斌’s key attribution is that the largest common obstacle to being “useful” and “trustworthy” is still low IQ. Robot arms, cameras, and many body components are already usable, but models cannot reliably understand the environment, execute tasks, or handle changes.
The company therefore treats model capability as the top priority, with near-term commercialization in second place. The 2 cannot be separated, however: commercialization must generate real feedback that improves the model, rather than becoming a standalone revenue line unrelated to the general-purpose model.
20. “Embodied-Native” Rejects Simply Bolting Action onto an Existing VLM
The common industry approach is to download an open-source VLM, attach an Action module at the end, and train a VLA. 唐文斌 compares it to entering a sports academy only after completing 9 years of compulsory education: the body was not trained during childhood, so the ceiling is constrained; focusing only on sports also causes the model to lose its Chinese and math.
In model terms, robot data should enter from the beginning of VLM pretraining rather than appearing only after the VLM is complete. VLA training should also co-train grounding, VQA, and other multimodal capabilities, rather than optimize Action alone and allow existing capabilities to deteriorate materially.
He compares native multimodality with grafted multimodality: from-scratch native training costs more, but is more likely to produce a better capability structure. 原力灵机 wants to extend the same principle from VLMs to embodied-native VLM/VLA systems.
21. DM0 Puts Internet, Autonomous-Driving, and Robot Data into One Training Chain from Day One
唐文斌 divides physical-world data into 3 categories: internet multimodal and video data, autonomous-driving data, and robot-operation data. He believes all 3 should be jointly trained from day one into a physical-world-native VLM, rather than forming separate silos.
A robot’s Action is not limited to the hands; leg movement and navigation are also Action. Visual Language Navigation and VLA therefore share a common foundation. If the goal is a robot operating across indoor and outdoor environments, autonomous-driving data can strengthen environmental understanding and navigation, not just desktop tasks.
DM0 training has 3 stages: 原力灵机 and 阶跃 mix multimodal, robot, and some autonomous-driving data to jointly complete VLM pretraining; 原力灵机 then continues with VLA training and finally post-training.
阶跃, 千里, and 原力灵机 are “sister companies,” in 唐文斌’s description, with business cooperation among them. Robot data comes from 原力灵机, autonomous-driving data from 千里, and 阶跃 brings accumulated multimodal data; a joint base model could benefit all parties.
22. Real-Robot Data Remains the Main Ingredient; Simulation Works Only for Specific Tasks
原力灵机 uses both robot-centric and human-centric data. The former includes teleoperation, hand-guided operation, exoskeletons, and collection without a robot body; the latter includes data generated by wearable devices. The final result depends on the right mixture or “recipe.”
唐文斌 remains cautious about simulation for contact-rich manipulation. Rigid bodies are manageable, but non-rigid bodies are hard to reproduce faithfully; force, tactile, and other multimodal signals are also difficult to simulate, and the sim-to-real gap limits simulation’s help for real-operation models.
Simulation is more valuable for 2 types of problems: locomotion, where the robot mainly interacts with the ground and can learn center-of-gravity and motion control at scale; and higher-level task understanding, decomposition, and planning, which do not require exact replication of contact details.
He remains open to whether world-model-generated data can scale. Video, first-person data, simulation, and real-use flywheels could all be viable paths, but whether generation quality will genuinely help existing models remains a “big if.” For now, real-robot data remains the priority.
23. Embodied Systems May Require 3 Models Operating at Different Frequencies, Not 2 Layers
System 2 handles slow thinking, reasoning, task understanding, and decomposition, potentially running in the cloud at a lower frequency. This is precisely where LLMs and foundation multimodal models are already strong.
System 1 handles mid-frequency robot operation, with VLA potentially running at around 10 Hz. But fine movements such as rotation and adjustment are difficult to control at only 10 Hz, so a System 0 may be needed below it, processing force, touch, and rapid feedback at above 100 Hz.
What intermediate representations should pass between the 3 models, and how to make the big brain’s output actually improve the cerebellum’s performance? 唐文斌 says plainly, “We actually don’t know yet.” This is not a minor engineering-interface issue, but a core unresolved question in embodied-model architecture.
The program’s closing discussion linked the idea to Sharp’s 3-layer structure presented at CES: foundation models are good at high-level planning, while startups have more room to differentiate in the tightly coupled software-hardware System 1 and System 0.
24. Foundation-Model Companies May Not Easily Cover Embodied Startups
唐文斌 compares it to “watching Messi play soccer”: watching enough videos does not necessarily teach you how to play. Embodied capability is tightly tied to specific hardware, contact feedback, and control frequency, and cannot automatically be solved by a larger video model.
OpenAI, Google, and ByteDance have clear advantages in System 2, but System 1 and System 0 still require joint iteration between models and hardware. Large companies can of course invest in hardware and settings, but every company has limited resources and attention.
The more important issue is where the data comes from. Collection data alone is not enough; the truly valuable data comes from failures, human interventions, and task outcomes during scaled use. Getting that data requires closing the application loop first, something a standalone model-training effort cannot do.
The industry therefore faces a chicken-and-egg problem: robots are immature, so they cannot be used at scale; without scaled use, there is no usage data. Whoever first gets a limited-capability model running continuously in the right setting may be the one to build a hard-to-replicate data moat.
25. Evaluation Is Moving from Specialist to Generalist, While Zero-Shot Is Still Too Early
唐文斌 divides evaluation into 3 levels. A specialist model is trained on extensive data for one task and fine-tuned on top of a base model to do only that task; this represents the capability path most of the industry is currently pursuing.
A generalist or multi-task model is given data for roughly 20 tasks and prompted to complete all of them with the same model, testing generalization more directly. Zero-shot goes further, withholding even target-task data and relying on prompts to complete an unseen task.
原力灵机’s generalist success rate is currently only in the 30%-plus range. 唐文斌 believes the important shift in 2026 will be from specialists to generalists; zero-shot may still be out of reach this year, so generalist performance must first improve materially.
Because the sim-to-real gap is large, model changes ultimately still require large-scale real-robot testing. A benchmark is not merely an external ranking; it gives R&D a signal about which direction works and along which dimension progress is occurring.
26. A Good Setting Today Must Tolerate Errors, Time, Generalization, and Long Operating Hours
原力灵机 sets 4 “necessary but not sufficient” conditions for a setting: it must allow robot errors or provide a fallback; it cannot impose an excessively tight cycle time; it must involve enough generalization to justify a model rather than bespoke automation; and it must support long daily operating hours to underpin the return on investment.
Automotive production lines violate the first 2 conditions: they have little tolerance for errors and strict cycle-time requirements. 唐文斌 says the pick-and-place tasks imagined from outside become constrained on site by packaging, auxiliary materials, containers, positions, and process requirements; the real complexity is far greater than the promotional story.
Even though 原力灵机 has relationships across the auto supply chain and 千里 also has auto plants, the company still says no to automotive production lines. 唐文斌 stresses that “many judgments have to be made on site”; the fact that Tesla or Figure talks about auto factories does not make them a naturally ideal setting.
The program’s closing discussion preserved another view of the setting question: 贾鹏 puts more emphasis on end-to-end tasks completed entirely by robots, while 唐文斌 accepts human fallback. What they share is avoiding work with excessively fast cycle times.
27. Logistics Can Serve as a “Greenhouse” for Robots with Limited Intelligence
Logistics systems already combine human labor with various forms of automation, making them relatively easy places to introduce a new robot whose capabilities are incomplete. When the robot cannot complete a task, the dispatch system can route the item to a human station and keep the process closed-loop.
唐文斌 compares it to a Didi or Uber network in which autonomous vehicles and human drivers coexist: no individual machine has to cover everything, because the system absorbs failures through division of labor, remote intervention, and dynamic dispatch.
This “greenhouse” does not conceal capability; it creates the conditions for continuous use. If one failure causes a robot to stop permanently, it will never receive real feedback. Allowing it to keep trying inside the system lets failure data flow back into the training pipeline.
28. Continuous Use Requires Testable Operating Standards, Not More Demos
唐文斌’s on-air definition was a real production or service setting, more than 10 hours of operation per day, 2 consecutive months, and “scale.” Pressed on what scale means, he initially cited 100 units as a rough number.
By that standard, he believes the industry may have almost no mature examples. Many videos prove only that a robot succeeded once; they do not say whether that was 1 success selected from 30 failures or the model’s stable average performance.
原力灵机 does not plan to launch across many industries at once. It wants to close the loop in 2 or 3 settings; 100 or even 1,000 robots working continuously in one setting is closer to commercial deployment than a dozen shallow cases.
The company has set a 2026 target of “one setting, 1,000 robots,” and requires those units to meet the continuous-operation standard. 唐文斌 views 1,000 units as a quantitative threshold at which the data flywheel might start, not simply a shipment target.
29. Teleoperation Validates the Hardware Ceiling; Models Cannot Define a Body Out of Thin Air
唐文斌 rejects the simplified claim that “the model defines the hardware.” To determine whether a body can solve a task, first test whether a human can complete it through teleoperation; if the human brain cannot control that body to complete the task, no algorithm can overcome its limits in precision, payload, or configuration.
A fully standardized body can easily be either underkill or overkill: camera placement or end-effector dimensions may prevent it from reaching the target at all; even if it can perform the task, excessive design or excessive mobility may destroy its ROI competitiveness.
A high-center-of-gravity humanoid or wheeled dual-arm robot is unlikely to move quickly. 唐文斌 estimates that 0.8-1 meter per second may already be close to the limit, with emergency stops creating a risk of tipping; a low-center-of-gravity AGV, by contrast, may reach 4 meters per second. The AGV can therefore move quickly while the manipulation robot stays at the workstation.
原力灵机’s formulation is: “The model unlocks the setting; we define the hardware through the setting.” The company will retain standard modules and combine them rapidly, Lego-style, into different forms rather than choosing between a standard humanoid and project-by-project customization.
30. Infrastructure Matters Because It Improves Iteration Speed, Not Because It Decorates an Ecosystem
唐文斌 says current model performance is still “the 50-step man laughing at the 100-step man”: leadership changes hands, and the real competitive edge is iteration speed. That depends on the full infrastructure stack—data, training, inference, hardware, and evaluation—not on a temporary leaderboard position.
DexBotic was initially built to reproduce algorithms internally, compare experiments, and unify data workflows. It is not a replacement for PyTorch, but a robotics-oriented toolbox on top of it, with configurable VLM base models, vision encoders, Action modules, diffusion paths, and deployment methods.
RoboChallenge uses large-scale real-robot testing to provide research signals, while DeepFlow targets workflows. 唐文斌’s logic is that without a reliable benchmark, the company cannot tell whether it is improving, and the industry cannot establish a common technical coordinate system.
He describes open infrastructure as “both altruistic and self-serving.” The industry’s most important question today is not which company takes the lead first, but “how fast this industry can actually work.” If models never reach the application threshold, every company may become a bubble outside research and entertainment.
31. MegEngine’s Lesson Is That Open Source Must Happen Before the Ecosystem Hardens
Megvii began building the deep-learning framework later known as MegEngine in 2013, when Caffe and MXNet were the only major options and TensorFlow had not yet emerged. The team treated it as an internal productivity tool and did not open-source it until 2018, by which point the ecosystem window had closed.
唐文斌’s conclusion is not that small companies should not build frameworks, but that “the key is to move early.” DexBotic is being opened while embodied development still lacks a dominant tool, allowing external use, modification, and contributions to improve internal efficiency in return.
DexBotic 1.0 was open-sourced in October of the previous year. 唐文斌 says it now has roughly 1,000 external developers and is used by a number of universities; the Qwen team has also exchanged ideas with 原力灵机 and assigned staff to develop on it.
Beyond imitation learning, the team found that RLinf had already built a strong reinforcement-learning framework, so it chose to connect interfaces and explore integration rather than duplicate the effort. Here, open source is first a trust mechanism: “The code is all in your hands,” allowing users to modify it themselves instead of being locked into a vendor.
32. Fairness in a Self-Run Leaderboard Depends on Full Reproducibility
DM0 ranked first on both RoboChallenge leaderboards, prompting questions about “being both athlete and referee.” 唐文斌 says the team debated the issue internally, but ultimately decided that whether it submitted was “not fundamental,” because the tests, code, and training details were public.
External teams can download DexBotic, the DM0 code, and the fine-tuning and generalist details for each task, then reproduce the scores directly. Fairness depends not on 原力灵机’s verbal assurance, but on whether anyone can rerun the results.
He stresses that RoboChallenge was not created to produce a permanent number-one ranking, but to establish a benchmark for measuring the industry’s real technical stage. Another company surpassing DM0 2 weeks later would be perfectly normal; the leaderboard should pull the field forward, not serve as a company endorsement.
33. The Final Contest Will Be Decided by Whether General-Purpose Models, Setting Flywheels, and Sustained Focus Can Become One System
唐文斌 expects many vertical companies to exist in the industry’s early phase. As large-model capabilities improve and smaller models are absorbed, capability will concentrate, potentially leaving only a single-digit number of global platform companies. He further expects most may come from China because embodied intelligence depends more heavily than pure foundation models on hardware supply chains and rapid body iteration, while acknowledging that Silicon Valley companies could still secure positions with models at their core.
To survive to the end, a company needs 2 things: a lead in model IQ and a data flywheel built through long-term robot use in real settings. The former is the soul; the latter continually feeds failures, interventions, and feedback back into the model. 唐文斌 still sees “one setting, 1,000 robots” as the flywheel’s starting point.
He agrees that going too deep into deployment can drag a company into single-setting optimization, so the model brought to market should point directly toward generality, with setting selection and systems engineering filling the gaps, rather than training an overfit model for every customer. “You have to be careful not to become a little-old-man company”—one that lives a long time, remains small, and is fundamentally still a bespoke project vendor.
Megvii’s biggest strategic regret is a lack of focus: the company kept adding vision, security, autonomous driving, robotics, and mobile initiatives, diluting even exceptionally strong talent. “A plus B plus C is basically equal to max(A, B, C).” If he could start over, 唐文斌 would concentrate more people on 1 or 2 things he truly believed in.
That is why 原力灵机 rejects showpiece robots, gala-style exposure, and the temptation of automotive production lines. In anticipation of a possible Gartner trough, its strategy is to “stockpile enough provisions,” building cash reserves when sentiment is strong without spending them through shotgun expansion. The stance is to be “neither too optimistic in the short term nor too pessimistic in the long term.”
The 2 hard targets for 2026 are to raise the current 30%-plus generalist success rate to an undisclosed but “extremely aggressive” level, and to have 1,000 robots running continuously in one setting. Unresolved questions include the interfaces between the 3 model layers, the most scalable data path, and whether a new model architecture is required.
唐文斌 describes his role as an “AI technology product manager”: someone who understands model boundaries and can use workarounds, workflows, and infrastructure to place technology precisely into customer settings. After 15 years of entrepreneurship, he still remembers the robot deliberately placed at Megvii’s front desk and calls embodied intelligence “worth spending a lifetime and all one’s time on.”