Yang Shuo: MiaoDong Technology, Tesla Optimus, CMU and DJI, Drones, Humanoid Robots | WhynotTV Podcast #1
Summary
Yang Shuo places today’s robots at the stage cars were in during the 1880s: products with real practical value may emerge over the next 1-2 years, human-sized systems capable of several household chores in 3-5 years, but arbitrary tasks such as “clearing the table after hot pot”（吃完火锅收桌子）will remain unsolved for 5-10 years. Commercialization will not wait for a sudden leap to general intelligence, but will expand task by task from moving boxes and picking up objects from the floor; “Product A, Product B, Product C” will take turns advancing through breakthroughs in mechanics, hardware, and algorithms.
What is truly difficult to replicate is not a single robot demo, but the full loop from technology and mass production to after-sales service and repairs; Yang Shuo therefore sees Tesla as currently closest to a “Boeing-level” reliability system. If Optimus really produces 5,000 units this year, Tesla’s existing infrastructure for deploying, testing, and analyzing failures in cars can absorb the data; MiaoDong Technology is also targeting one-in-a-million reliability because “there is risk in doing this, and we may fail, but it is still better than doing nothing.”
MiaoDong Technology is pursuing a product strategy tilted toward To C, pairing Shenzhen’s supply chain with Silicon Valley’s innovation culture in a bi-coastal team, prioritizing validation of hardware mass production, after-sales service, and repairs while expanding intelligence through a more open ecosystem strategy. Yang Shuo believes VLA and hardware could each consume an entire company’s resources, and few companies funded purely through financial investment can do both well at once; industrial specialization and collaboration are more sustainable than building everything in-house.
Humanoid robots will not unify every use case; they are simply powerful “Swiss Army knives,” and homes may ultimately accommodate wheeled robots, humanoids, and fixed robotic arms at the same time. A wheeled robot could cook in the kitchen, a humanoid could carry things upstairs, and a ceiling-mounted arm could work in the laundry room; investment and product decisions should work backward from user problems to form factor, rather than “looking for a nail with a hammer.”
Once the technology matures, Yang Shuo expects the robotics market to support 10-15 large companies and reach a scale of hundreds of billions to trillions, potentially producing companies as large as Apple or Nvidia. Competition already spans capital, hiring, suppliers, and model technology, but technological convergence will not erase differentiation: the real moat is the long-term accumulation of details across training, deployment, software, hardware, and human-machine interaction, and “those details, piled up together, create enormous differences and moats.”
Near-term progress involves both object representations in neural networks and unglamorous but foundational technologies—motor overheating, materials, and thermal management—that determine operating conditions. Yang Shuo even imagines outdoor robots “taking up water to drink” 5-10 years from now, using evaporation to cool their motors; he also admits it was a regret not to have studied reinforcement learning earlier, while arguing that MPC and control theory will continue to find applications in reinforcement learning through simulation and optimization.
Yang Shuo sees robots as the continuation of human tool evolution, with the long-term keywords being not replacement but “coexistence” and “death.” In his projection, digital life capable of copying memories and weights may need to set an endpoint at 50 or 80 years because “death is actually a source of creativity”; brought back to the present, his life reward function contains only the happiness of his family and children, and building MiaoDong’s products well.
Deep dive
1. A bi-coastal startup connects Silicon Valley’s appetite for risk to Shenzhen’s supply chain
Yang Shuo began working on DJI’s underlying flight-control systems as a graduate student in 2013, rebuilding its early algorithms and software stack; from 2015 to 2018, he worked full-time on product projects and was responsible for RoboMaster.
He went to CMU in 2018 for a PhD in robotics, joined Tesla Optimus in 2023, spent a year and a half there, and then founded his own company; the experience across China and the US convinced him that engineers and industrial ecosystems in the two places each have distinct strengths, making a bi-coastal team necessary.
His trade-off is clear: Silicon Valley still encourages failure, innovation, and talent mobility, while China has a significant supply-chain advantage; MiaoDong aims to “bring people with different backgrounds and different ideas together into one team.”
2. Drones took years to go from “able to fly” to one-in-a-million reliability
He Tairan recalled DJI’s 2016 whiteboard: Boeing was the biggest name among the competitors, with a failure rate of roughly one in five million, while the Phantom 4 was around one in two million.
Yang Shuo used the 2009 quadcopter scene in 3 Idiots to illustrate that drones had already been developing for nearly a decade before 2009; the Phantom 1 launched in 2013 and reached meaningful sales, but it took several more years of large-scale use before reliability became a core metric.
In his view, humanoid robots have not even reached drones’ 2009 stage, with many basic technologies still closer to research; but existing experience in electromechanical systems means “there should be” some commercial consumer robots within 2 years.
MiaoDong is still targeting one-in-a-million reliability. Yang Shuo acknowledges that aiming for mass production carries risk and may fail, but if nobody works through the details one by one, “the future that can be mass-produced and stable will definitely not arrive.”
3. Tesla’s lead lies not only in Optimus itself, but in the infrastructure inherited from cars
Asked which company is closest to Boeing-level reliability, Yang Shuo chose Tesla: robot deployment, hardware, data collection, and testing can all inherit systems already built by its automotive business.
His conditional statement is worth preserving: if Tesla really builds 5,000 Optimus units this year, there will already be processes to handle failure collection and analysis wherever they are deployed; most startups have never operated at that scale.
Solid technology is only the first step. Mass production, after-sales service, and repairs come next: when performance degrades, does the unit go back to a repair center or get shipped to the company’s service workshop? These “tiny, detailed problems” collectively determine whether a hardware product can work.
4. Good products start from the user’s problem, not from finding jobs for humanoids
On whether DJI will build a humanoid robot, Yang Shuo admitted that he has been away for years and does not know its strategy; based on product logic alone, he expects DJI would first study how to continue freeing image creators’ productivity and creativity.
If users need someone to carry an Osmo and follow them while filming, the answer could be a humanoid robot, but it could just as well be a small car, a lower-body platform, or a wearable stabilizing device similar to Ronin; if the final experience is the same, none of these solutions has an inherent advantage.
To him, the humanoid is a highly capable hammer, even a “Swiss Army knife,” but many scenarios need only an ordinary hammer. That does not mean the bubble must burst; it means hardware is expensive and software is not yet smart enough, while general-purpose tools will not eliminate specialized ones.
5. This humanoid cycle will be bigger, but it remains subject to the industrial cycle
Looking back over the past 50 years, Yang Shuo sees Japanese concepts in the 1980s, ASIMO in 2000, and the small consumer humanoids that failed afterward as evidence that the industry has always moved “in waves.”
He is certain that today’s technology is ahead of the previous generation and that this cycle can at least achieve more, but he does not believe it is sufficient to cover every imagined household scenario: “We can do a little better than they did, but we are still far from our own dream.”
Progress will not arrive as one system that suddenly surpasses humans. It will accumulate across Product A, B, and C: mechanical innovation today, hardware tomorrow, and algorithms later.
6. Robots “drinking water” exposes the materials bottlenecks hidden by demo videos
Existing motors cannot operate reliably across all conditions; whether Unitree or Tesla robots, sustained outdoor work in hot weather creates overheating problems that directly limit the application envelope.
Yang Shuo imagines new conductor materials, redesigned motors, or a robot carrying its own water tank and delivering water to the motor for cooling through a process similar to sweating and evaporation.
He therefore says he would not be surprised to see an outdoor robot “take up water to drink” 5 or 10 years from now. Cars need coolant, and robots cannot escape the physics of heat generation.
Product iteration means first finding scenarios that work, then expanding operating conditions with new materials, hardware, and algorithms, rather than waiting for a universal technology package to arrive all at once.
7. Yang Shuo took back his early view that “quadrotors should no longer be studied”
Advising students in 2018 not to study quadrotors was, Yang Shuo now clearly admits, a mistake: he confused whether the engineering had been fully worked through with whether the platform still had research value.
The more mature quadrotor hardware and low-level control become, the more researchers can focus on navigation and path planning; its small size, high speed, and maneuverability instead make it an ideal test platform.
Borrowing von Kármán’s distinction, he sees a scientist as someone who “explains what already exists,” while an engineer “creates what does not exist.” A robotics PhD often has to build the platform first, then explain why method A is inferior to method B and what the network architecture and training method actually contribute.
8. The gap between papers and products is mainly systems engineering, not one more algorithm
Once research enters a product, it must simultaneously fit technical implementation, mass production, after-sales service, and repairs; any link can generate technical or scientific problems, and the work depends more on continuously collecting failure data and solving small issues.
Because robotics as a whole is still immature, research and engineering will alternate over short time horizons; but “building it” and “explaining it clearly” remain different kinds of work, and one demo cannot substitute for product validation.
9. He regrets missing reinforcement learning, but MPC was not wasted effort
Looking back at his choice between CMU and Berkeley, Yang Shuo said plainly that it was “definitely wrong”: if he could do it again, he might study reinforcement learning directly and earlier, while acknowledging that he did not understand model-training details deeply enough.
Model-based methods, especially MPC and optimization, will still live inside reinforcement learning. A lesson from Zhang Wei at SUSTech was that the accumulated knowledge of traditional control was not wasted; it simply reappears through another layer.
Physics simulation on GPUs is itself an “extremely model-based” optimization system, still dealing with kinematics, numerical integration, estimation, and contact constraints; human understanding of the model is embedded in the simulator and then implicitly learned by the neural network.
10. The next generation still needs signal systems, but not every classic derivation
Yang Shuo does not advise abandoning fundamentals, but sees no need to master the 4 derivations of the Kalman filter; more important is understanding its links to basic mathematics, signal processing, and frequency superposition.
Faced with the bottleneck of being able to fold one red shirt but not all clothes, the highest-value area for investment is object representation: abstracting low-dimensional structure from sensor, image, and robot data so that the structure genuinely helps execute tasks.
11. Musk applies pressure to Optimus, but remains patient with frontier uncertainty
Yang Shuo described Elon Musk in work settings as “very professional”: he watches the speaker and the slides carefully, does not get distracted, and rarely asks inefficient questions.
A meeting lasting roughly 1.5 hours might hear 20-30 frontline engineers speak for a few minutes each; Musk directly collects technical and business updates, then asks project leads to adjust direction at a high level.
The Optimus team did not experience the “being fired mid-presentation” incident Yang Shuo had personally witnessed. He had heard that it could happen on other make-or-break projects, but on humanoids, Musk knows every detail is at the frontier and carries limited certainty.
12. Mission turns new technology into a new category, while hardware is beginning to take shape
Yang Shuo sees a common thread between DJI and Tesla: both insist on creating new categories and presenting near-mature technology through entirely new experiences; he also directly agreed that this sense of mission helped drive both companies’ success.
Revisiting He Tairan’s Michelangelo analogy, Yang Shuo believes the hardware system of a human-sized robot has “almost carved a usable hardware system out of metal”: basic walking, teleoperated clothes folding, and autonomous object pickup have all shown signs of progress.
He expects complete systems capable of useful household chores within 3-5 years; expanding from one task to 50 or 100 could take another 5 or 10 years, with the constant prerequisite that intelligence in a given scenario be “useful enough.”
13. Household robots are more likely to be a coordinated team than one general-purpose servant
Humanoids, wheeled robots, drones, and ceiling-mounted robotic arms will eventually become ordinary options in the toolbox; humanoids are the coolest today, but will not retain a permanent special status.
In Yang Shuo’s imagined home, a wheeled robot cooks in the kitchen, a humanoid carries things upstairs, and a ceiling-mounted arm in the laundry room loads the washing machine, removes the clothes, and folds them, without even needing to move outside the room.
Humans already divide labor among live-in help, hourly workers, and chefs, while kitchens accommodate both microwaves and dishwashers; what robotics companies really need to do is “combine different tools to solve people’s problems.”
14. Commercialization starts with moving boxes and picking objects off the floor, not clearing the hot-pot table
Grabbing a beverage bottle is different from grabbing Lego; Yang Shuo noted that Agility has demonstrated robots moving boxes continuously for 6-7 hours without failure, putting these structured logistics tasks close to commercial deployment.
He Tairan’s counterpoint was that replacing car parts with Amazon packages still resembles assembly-line picking from decades ago, while users really want to say, “I just had hot pot—clear the table.” Yang Shuo’s answer was, “I don’t think so,” and he does not expect that problem to be solved in one step within 5-10 years.
A more realistic product would add an arm to a robot vacuum, allowing it to collect objects within reach into a fixed bin, with commercial trade-offs keeping the price at RMB1,000-2,000; it is not sexy, but could unlock real willingness to pay.
Yang Shuo compares today with the early automobile industry of the 1880s: cars already existed in the 1880s, after 40-50 years of design iteration. Robots do not need to wait another 100 years, but they are still in the early stage dictated by objective laws.
15. MiaoDong is betting on a To C loop while keeping intelligence open to collaboration
Hardware and foundation models have both begun to show promise, but having one company push both forward like Tesla requires enormous capital, people, and time; it is equally necessary for different teams to explore separately and share resources.
MiaoDong is more To C-oriented, prioritizing tests of hardware mass production, after-sales service, and repairs while expanding intelligence; Yang Shuo wants a more open, ecosystem-based strategy rather than building every model capability internally.
His resource assessment is direct: VLA may require investment comparable to the entire hardware stack, and few companies funded purely from a financial-investment perspective can do both well at once, making it easy to “get both things only 60% done”; it is better for different companies to form an industrial collaboration.
16. Technological convergence is not the danger; betting on the wrong shared path is the industry’s true tail risk
Competition among robotics startups now spans fundraising, hiring, suppliers, and model technology; Yang Shuo sees this as positive because competition trains talent, drives component mass production, and forces teams to keep moving quickly.
Once household robots become indispensable, he expects 10-15 large companies globally and a market worth hundreds of billions to trillions; as with different automobile categories, different robot companies will serve different customers through their respective strengths, and robot companies the size of Apple or Nvidia may emerge.
He Tairan worries that hardware and VLA architectures are becoming increasingly similar, and that if everyone follows the same wrong skill tree, this wave could “end up very badly.” Yang Shuo acknowledges the possibility, but points out that assembly-line-style breakthroughs spread, while knowledge and code in VLA and reinforcement learning are also widely shared.
In an ecosystem with software and hardware divided, software teams can change direction and hardware companies can switch partners; he therefore attributes the ability to follow breakthroughs more to organizational management than to a single technology that remains closed forever.
17. The moat is the compounding of countless details; investors must let long-term technology fund short-term business
Deep learning continues to disrupt old technologies, while academic sharing shortens replication time; Yang Shuo believes the only moat comes from continuously refining details across model training, deployment, software, hardware, and human-machine interaction.
Twelve years after DJI launched the Phantom line, flight control, cameras, and transmission solutions had long since matured, yet products assembled by other manufacturers still lagged DJI; robots have more motors and more varied tasks, so the required detail is “not just one order of magnitude greater.”
If he were an investor, Yang Shuo would look at foundational breakthroughs that may not enter humanoid robotics within 5 years, including motor materials, conductors, and evaporative cooling, while allowing teams to generate revenue first by selling joint motors or other existing motor products.
A foundation model could “burn through $100M in a year,” so a long-term mission must rest on the ability to generate cash. MiaoDong has likewise chosen To C based on its own experience, using short-, long-, and longer-term plans to demonstrate the commercial loop.
18. Human-machine coexistence will eventually make “whether machines should die” a social question
Yang Shuo describes the history of tools as an extension of biological evolution: stone and wood combined into an axe, followed by the wheel, plow, clock, typewriter, car, airplane, and robot; human capability has always been defined in part by the improvement of machine capability.
In a more distant projection, humans may gradually “de-embody,” moving consciousness into mechanical bodies. Using the contrast between hair in a bathtub, which disgusts people, and electrical wires, which do not, he says “human development is actually the continual elimination of our own bodies,” but places that scenario thousands of years out.
He is not worried about Terminator-style replacement: humans today cannot live without machines, and conscious robots in the future may not agree to destroy humanity, just as humans would not agree to smash every refrigerator, television, and washing machine; the core keyword is “coexistence.”
The sharper question is whether digital life capable of copying memory, data, and weights—approaching immortality—should be subject to a 50- or 80-year limit in a machine society, with backups prohibited. Yang Shuo’s reason is that “death is actually a source of creativity”: limited time creates action, remembrance, and new emotional experiences.
19. The grand mission ultimately compresses into family and product
He Tairan spoke of developing feelings for H1 when the experiment ended, and Yang Shuo admitted that leaving Optimus made him uncomfortable for days; even a clumsy robot can make parting a real experience after sustained daily interaction.
Yang Shuo defines his role as helping robots grow, helping human friends understand how to coexist with robots, and eventually helping future robots learn to coexist with humans; founding a company or doing research is simply where the mission lands at different stages.
His current reward function is deliberately narrow: first take care of his children and his family’s happiness, then build MiaoDong’s robot products well. “In the same period of time, you can only do 2 or 3 things well.”
He Tairan’s final takeaway was that the directions and skill trees that seem important in 2025 may all be overturned 10 years from now, but the belief that “Robotics matters, and I’m willing to solve this problem” will not be overturned; Yang Shuo agreed that this is more durable than chasing any single technology.