141: Strutt’s 洪小平: Former DJI LiDAR Head Builds an “Electric Wheelchair”? Embodied Intelligence Without a Humanoid
Summary
洪小平’s core view is that the “Everest” of embodied intelligence is clearly visible, but the path of building humanoid robots directly “doesn’t appear viable at this point.” Language models train on roughly 1T tokens as a one-dimensional time series, while robots must handle at least five dimensions—time, 3D space, touch, sound and more—with data-coverage difficulty rising sharply with each added dimension. He acknowledges that logic encoded in language and correlations across modalities may reduce the requirement, but “a few hundred hours of video data” is nowhere near enough to train a reliable general-purpose robot, and no source yet continuously generates household-life data at scale.
Strutt is addressing aging and personal mobility first so a sellable product can “lay eggs along the way” toward long-term embodied intelligence. High-end senior communities may charge $500K–$1M to enter, while airports in the US and Singapore struggle to find people to push wheelchairs. 洪小平’s goal is therefore to “solve real problems for a lot of people today,” using real-world usage and human-life data to drive iteration instead of waiting 10 years for a technological holy grail.
EV One is not a conventional electric wheelchair; it is an Everyday Vehicle filling the gap between walking distance and walking speed. Its users include not only people with disabilities, but also those who walk with pain, have balance issues, tire easily, or simply want to walk less at airports, trade shows and parks. The product deliberately uses a light, sporty design to shed the medical stigma of wheelchairs—and the assumption that someone who uses one must also have cognitive problems.
EV One’s product value lies in EV Sense, Quad Drive and low-friction human-machine interaction—not in piling on humanoid aesthetics. EV Sense includes obstacle-aware slowing and stopping, button-authorized steering, two types of Copilot, Waypoint for fixed destinations and Path Finder for temporary targets, plus two types of Autopilot. Four-wheel drive, suspension, self-stabilization and ABS serve safety and comfort; a large language model turns “I’m thirsty” into actions such as going to the refrigerator, while creating an expansion point for future robotic arms and accessories.
For the first generation, the execution moat is production reliability, not prototype demos. From summer 2023 to summer 2024, the team developed the chassis, four-wheel drive, sensor fusion and autonomous-driving stack before cutting height adjustment, forward-tilt and other reconfigurable mechanisms. Reliability testing delayed the product by nearly 6 months; the dual-drum test with 5 cm bumps raised the target from the usual roughly 200,000 cycles to 2M, because “building a prototype is easy, but there are many, many pitfalls between that and mass production.”
The strongest proof of demand is users regaining life experiences—not a preorder figure. A Florida high-school graduate with a fragile neck had once crashed at home in a rented power wheelchair. Years after she stopped visiting a muddy stable, EV One’s suspension and driving capability let her return to brush her horse; her mother burst into tears. The case showed the team how mobility products can help older people, young people, people with disabilities and people without them.
The biggest commercialization unknown is where the new category ends; pricing and the regulatory path remain unsettled. The team leans premium but does not want the product to become unaffordable, with survey responses ranging from $1,000 to $20,000. The first product will be sold as personal mobility equipment and may not qualify for insurance reimbursement; dedicated SKUs for reimbursed or medical-device settings could come later. 洪小平’s view is that many supposed “no-market” problems actually mean P is not good enough, not that M does not exist.
A first-mover brand, patents and a complementary team can buy time, but cannot replace iteration or crossing the chasm. The first credible brand in a safety product can establish the category in consumers’ minds, while joystick authorization, four-wheel drive and suspension, mode switching and language-model integration have all been patented. But 洪小平 says the company cannot stay ahead for 5–10 years with one product; he is still testing EV One’s true niche market. Even if a smaller chasm brings red-ocean competition faster, he would still choose that “difficulty within optimism.”
Deep dive
1. 洪小平 moved from “doing research” to putting technology into everyday life
洪小平 studied the optical properties of nanomaterials at Berkeley. His academic record was strong enough that many people expected him to remain in academia, and he had published in Nature and papers including in Science. But 1–2 years before finishing his PhD, he realized that although this “may have achieved some of my ideals and pursuits in life,” it was not the life he wanted most.
He eventually settled on combining technology and the humanities: engineering products, refining them to the finish line, and connecting technology with ordinary people’s daily lives. Fresh out of his PhD, he knew how to conduct research but not how to turn a concept into a product that could be mass-produced.
After spending 1–2 years at Honeywell working on gas sensors, he was invited back to China by his undergraduate classmate Wang Mingyu, then DJI’s VP of R&D. 洪小平 “wanted to start a company but didn’t know how to begin,” so he viewed DJI as a platform for entering an early-stage company and learning the full product chain.
2. Livox’s low price came from designing for mass production from day one
DJI had already been around for 10 years in 2016, but was still competing with Intel, 3D Robotics and rivals backed by domestic capital. The company believed drones needed intelligence to stand out, and intelligence required a new generation of sensors.
The optoelectronics team led by 洪小平 initially developed LiDAR for drones and later served the automotive market, including the XPeng P5. Its first-principles approach was not to build a prototype first and hope to reduce costs later, but that “every design actually needs to serve mass production.”
Components had to be purchasable at the time, and manufacturing processes had to be ready to use—or at least demonstrably feasible from first principles. Livox launched its first product at CES 2020 for under $1,000 and was dubbed a “price killer.” 洪小平 saw the low price simply as the result of designing for mass production.
3. Leaving DJI was not a return to academia, but a way to fill gaps in products and social insight
When 洪小平 left DJI around 2019, he had learned how to take a product “from concept all the way to mass production,” but LiDAR was still only a sensor serving an end product. What he really wanted was to build an end product that “directly serves people and humanity.”
He then joined a new engineering program at SUSTech run by Wu Jingshen and Li Zexiang, teaching electronics and programming for roughly 3 years. The program trained students through projects, hands-on work and social problems. It also showed him that exam-oriented engineering education often leaves engineers trapped in test-taking logic when they face real-world needs.
Looking back, he admits that “a little less agonizing and a little more impulse” might have led him to start a company earlier. But technical accumulation, understanding human nature and a sense of social responsibility all take time. He believes waiting a few more years could have cost him the opportunity, while starting now is not too late.
4. PaLM-E and RT-2 revealed the robot’s “higher-level thinking”
洪小平 had viewed robotics as the direction he wanted to pursue since 2019, but had no specific product in mind. He spoke with Wei Jidong of SLAMTEC Robotics and kept watching which application paths were clear and which remained concepts.
What finally pushed him to act was a series of Google projects from late 2022 to early 2023, including PaLM, PaLM-E, SayCan and RT-2, along with the VLA concept they introduced. Once language models were fused with sensors, robots were no longer limited to finite model-based actions; they began to show higher-level understanding. Chain of Thought was also discussed in related papers.
His interim conclusion became: “The past 10 years were robotics’ silver age; the next 10 will be its golden age.” If he did not start a company now, he worried he would miss the window.
5. Humanoids are the Everest of embodied intelligence, but there is no verifiable route to the summit
When investors repeatedly asked why he was not building a humanoid, 洪小平 answered that the team had indeed identified an Everest—or holy grail—but that the current humanoid path “doesn’t appear viable.” The core bottleneck is that data and algorithms remain far from sufficient to support embodied intelligence.
His order-of-magnitude comparison is that models such as ChatGPT and DeepSeek use roughly 1T tokens, effectively covering humanity’s accumulated knowledge, but primarily process one-dimensional language sequences. Robots must additionally cover 3D space, time, touch, sound and even smell; increasing dimensionality typically makes the required data coverage grow exponentially.
He retains an important hedge: language already encodes substantial logic, and the different perceptual dimensions are highly correlated, so robots may not need a simple exponential increase in data. But regardless, “a few hundred hours of video data today” is nowhere near enough, and he has not seen a setting capable of continuously generating massive amounts of household-life data.
6. OpenAI’s scaling success cannot be directly applied to the physical world
Cheng Manqi asked whether Chinese companies’ emphasis on “laying eggs along the way” was simply a response to insufficient resources, while OpenAI had pursued a high target from the outset. 洪小平’s distinction was that OpenAI had already seen early signs, but lacked the resources to pursue them and needed to concentrate capital and talent. Transformer, available data and a series of algorithms already existed.
OpenAI may have judged that Google was not paying particular attention internally, then used steadily larger networks and more data to validate scaling. Once large models reached a certain level of network and data scale, intelligence began to emerge, with some capabilities surpassing human cognition; capital then flooded in.
On whether building concrete products could cause a company to miss early signals, he believes research and data are highly public: “Everyone will see it when the time comes.” The real difference is whether a company recognizes the importance of the signal. Frontier breakthroughs may first emerge in universities, research institutes or Google-like organizations in a “scattershot” fashion, after which companies decide whether to place a major bet.
7. Household robots are trapped in a chicken-and-egg loop by the high reliability threshold
Factory tasks are controlled and data can be collected quickly. A robot entering the home, however, must understand how people move and what help they need in real life—which requires the product to enter homes first.
洪小平 believes a consumer robot will fail if it can only complete 50% of general-purpose tasks. The other half is enough to leave the home “a complete mess, with chaos everywhere.” That creates a closed loop: without data, reliability cannot reach the required level; without reliability, nobody buys the product, so there is still no data.
On the 1X NEO Gamma, priced at roughly $20,000 and still requiring human teleoperation for complex tasks, 洪小平 was blunt: “Pretty ridiculous,” and “completely useless.” Users will not pay for an expensive machine that still needs a person to control it.
8. Aging combines a social problem, paying demand and a data gateway
The structural trend 洪小平 sees is that wealthier regions age more severely, have fewer young people and find it increasingly difficult to recruit caregivers. Some high-end senior communities in the US and parts of Japan require $500K–$1M to enter, reflecting the scarcity of care workers.
Airports face the same labor shortage. The US, Singapore and other markets struggle to hire wheelchair attendants, while airlines still need to guarantee equal access to air travel. Assisted mobility is therefore a real social problem.
His startup priority follows from that: if robots can help people, they should first serve “those who need it most.” Every real-world use solves an immediate problem while generating data from interaction with people in real-life settings—the exact asset humanoid robots struggle most to obtain today.
9. Mobility is the first “egg” robotics can deliver immediately
The team wanted a product that could live alongside people, operate indoors and outdoors, and solve a clearly defined problem. It ultimately identified mobility as the most fundamental need. Airports, trade shows, amusement parks and public parks all contain a “last mile” for people: users may be able to walk, but they may be in pain, have balance issues, tire easily or simply not want to spend all their energy getting there.
洪小平’s ideal is simple: the device travels with the user and helps them move; it can carry luggage, food, water and medication, and follow the user when needed. That creates a useful service relationship earlier than an abstract general-purpose robot could.
10. EV Sense allocates control through two Copilots and two Autopilots
Basic Copilot precisely slows the vehicle before an obstacle and brings it to a stop. Copilot Plus is for hard-to-control situations such as narrow passages, blind spots while reversing and elevators; once authorized, it makes measured steering adjustments based on the environment.
The two Autopilots serve different objectives. Waypoint lets users tell the vehicle to go to places such as the refrigerator or bathroom, while Path Finder lets users tap a temporary destination on a map without naming it. Follow mode keeps the vehicle alongside the user when they walk and nearby when they need to rest.
The team initially planned to focus only on Copilot, but EV Sense already contained perception, planning and navigation at the base layer. When users requested Autopilot, the team could extend the system directly. 洪小平 sees that upgradeability as a platform strategy rather than a one-time feature list.
11. Quad Drive brings automotive chassis logic to the worst sidewalks
Quad Drive uses four-wheel drive to handle the slopes of San Francisco and Chongqing, along with grass, exposed roots and mud. 洪小平 is explicit that beaches “may not work that well,” rather than packaging the product as unconditionally all-terrain.
The vehicle also includes self-balancing, self-stabilization, ABS, slip detection and suspension. It must climb slopes while reducing vibration over rough surfaces. The chassis is not about showing off specifications; it is about making users feel “comfortable, safe and at ease.”
洪小平 emphasizes that sidewalks are often worse than roads: uneven brickwork, cracks and debris go unrepaired for years, and sidewalks may be blocked altogether. Automotive chassis have already validated the combination of power, suspension and stability. EV One is applying that logic to sidewalks.
12. Removing the “wheelchair feel” is itself a product feature
EV One looks more like an office chair mounted on a mobile chassis: light and sporty rather than like a conventional medical device. Many older people do not consider themselves disabled; they simply cannot walk far and tire easily. The wheelchair label can block adoption outright.
Disabled users described an even harsher stigma. One person said that when buying coffee, the staff would not ask him directly but would turn to the caregiver beside him—as if seeing someone in a wheelchair meant assuming that “his brain had a problem too.” Strutt wants to avoid reproducing that identity signal.
Cheng Manqi asked how the team could make trade-offs after defining the user as “everyone.” 洪小平’s answer was not to add a feature for every diagnosis, but to focus on shared needs: a comfortable seat, strong surface adaptability, simple controls and sufficient intelligence.
“For everyone” therefore does not mean satisfying every special requirement in generation one. It means first making safety, comfort and ease of use exceptional; different SKUs for reimbursement or medical-device settings may come later.
13. Existing power wheelchairs meet basic needs but do not expand people’s radius of life
洪小平 uses his grandmother, who is in her 80s, as an example. She moved from a bicycle to a more stable tricycle, but now cannot ride the tricycle safely either. Older people in the US and Netherlands buying e-bikes show the same demand: they cannot walk far, cannot ride anymore, but do not want to use a wheelchair.
Conventional power wheelchairs have an obvious medical identity and are not easy to control, making complex surfaces difficult to navigate. Wheelchair makers generally meet basic living needs rather than helping users live better and go more places.
While in the US, 洪小平 once tried another mobility vehicle traversing a ramp diagonally. Its motor lacked torque; although he steered right, the vehicle tipped left and nearly rolled over. The experience made chassis adaptability and intelligence part of safety—not optional premium features.
14. A supposed lack of market may simply mean the product has not crossed the usability threshold
洪小平’s reflection on PMF is that founders often focus on M, but “sometimes P is good enough” to expand what initially looks like a narrow market. An unexpressed need does not mean no need exists.
He uses 3D printers as an example. Users once had to level the machine, assemble it, run repeated tests and wait hours or even tens of hours. The learning curve confined the market to hobbyists. Once the product simplified the process and the platform added model sharing, “even beginners could use it,” and the market boundary moved.
Lawn robots were similarly constrained by the complexity of installing buried boundary wires; boundary-free intelligent systems delivered a major product leap. But until they truly reach the point of “not needing people,” they remain in the early-adopter and innovator stage, perhaps edging into the late majority. Robot vacuums followed the same path to mass adoption.
15. The hardest first-generation decision was not what to add, but what to cut for mass production
From summer 2023 to summer 2024, the team did not immediately lock in the product form. It developed the chassis, four-wheel drive, autonomous driving and sensor fusion while starting user research from Day One. Only in summer 2024 did the product form converge on this direction.
Deleted designs included variable vehicle height, seat elevation and forward tilt. Forward tilt could help older people with weak leg strength get on and off, while a lower seat could support activities such as eating. But every additional motor and linkage increases complexity in reliability, consistency, maintenance and validation cycles.
洪小平 acknowledges that “conceptualizing a product is easy, and building a prototype is easy.” A startup does not have the resources to make every good idea reliable in generation one. Those features were therefore left for generation two or three instead of being forced into the first product.
The team’s non-negotiable priorities were intelligence, seat comfort and chassis capability, ultimately unified as safety, comfort and ease of use. EV was also reinterpreted as Everyday Vehicle: “usable every day, in every kind of situation.”
16. A hold-to-authorize button resolved the control conflict in automatic obstacle avoidance
At CES in early 2025, stopping and automatic avoidance were still separate modes. Real users almost never used avoidance: if the vehicle actively went around an obstacle, it could no longer approach the obstacle, while repeatedly switching modes—even by pressing a single button—felt burdensome.
The team ultimately turned Copilot Plus into a hold-to-authorize control on the joystick. Users hold it in a narrow passage, elevator or while reversing, and the vehicle temporarily corrects its heading; releasing it immediately returns to normal driving. Changing direction autonomously can create a sense of lost control, while the button creates the mental model that “I am authorizing it to help change my direction.”
In interviews, users said they wanted Copilot but not Autopilot. Their actual behavior was more nuanced: unfamiliar roads and outdoor settings required a sense of control, while at home or in a familiar supermarket they wanted to stop driving completely so they could look at their phones, chat or enjoy the view. The gap between survey answers and real-world use defined the two feature sets.
17. EV One’s end state is an extensible robotic helper
洪小平 was deeply influenced by Google’s Everyday Robots project in 2017. The project explored helper robots, first with one arm and later two, and became a prototype platform for work including SayCan, PaLM-E and RT-2. EV One does not copy the humanoid form; it first uses Autopilot to take people where they want to go.
A large language model lowers the interaction barrier. Users do not need to say “go to the refrigerator”; they can say, “I’m thirsty—I want a cold Coke.” After testing the vehicle, a senior-community user told the team, “I not only really enjoy this vehicle, I also really enjoy talking,” because caregivers often have no time to chat.
The vehicle reserves 200–500W of power output, control signals, high-speed networking, a rear rack, a left-side accessory mount, USB-C and ports beneath the seat. These can connect automatic charging docks, sensors, computing units, actuators and eventually robotic arms. That is why 洪小平 says: “It’s not a wheelchair…it’s a platform.”
A robotic arm remains a long-term direction; the current focus is mobility and mass production. The team will not promise every possible action from day one. It will first make a small number of defined scenarios reliable, then let tens of thousands or hundreds of thousands of devices try new tasks in homes and continue development based on user feedback. That is his vision for an embodied-data platform.
18. A 2M-cycle reliability test drew the line between prototype and product
Starting around August last year, the team built a CES-ready prototype in roughly 3 months. What truly slowed mass production was validation: the project was delayed by nearly half a year, largely for reliability testing rather than immediate tooling.
The dual-drum machine used drums with 5 cm bumps, jolting the vehicle once per rotation. 洪小平 says relevant safety standards typically call for roughly 200,000 cycles, while the team raised its target to 2M. Mechanical connections, high-power wiring, control signals and high-bandwidth transmission all had to remain stable because users will not accept a vehicle breaking down or coming apart en route.
19. A Florida girl’s return to the stable showed that mobility restores life experience
A Florida girl about to graduate high school and attend college has a fragile neck that cannot withstand hard jolts. She currently uses a manual wheelchair. After renting a power wheelchair and crashing at home on the first day, she carried a heavy psychological burden around vehicle control.
She had loved riding and interacting with horses, but had not visited a stable for years. EV One’s suspension carried her over mud, grass and exposed roots so she could stand beside her horse and brush it again. Her mother cried on the spot because her daughter could finally get close to someone she could confide in again.
The experience showed the team that robotics can help older people, young people, people with disabilities and people without them—bringing practical convenience and the joy of going places.
20. Pricing, regulation and market size remain the most honest commercialization unknowns
EV One plans to announce pricing at CES in January 2026. The current direction is premium, but the team “doesn’t want it to become unaffordable.” User surveys produced willingness-to-pay figures from $1,000 to $20,000, and the team has not decided where within that range to price it.
The first product will be defined as personal mobility equipment rather than a medical device. People with disabilities can still buy it, but it may not qualify for insurance reimbursement. 洪小平 leaves open the possibility of dedicated SKUs for reimbursed or medical-device settings, which would involve a different set of requirements and trade-offs.
Conventional power wheelchairs can sell for tens of thousands of dollars, but industry reports are not particularly reliable. 洪小平 quotes a colleague: “The industry is primitive enough that it doesn’t even have one decent market-research report, and new enough that nobody knows who this product’s audience will be.”
EV One overlaps with power wheelchairs but is not the same thing. 洪小平 sees the opportunity in the “transition from unknown to known,” rather than using an inaccurate market report to manufacture certainty around a new category.
21. A complementary team is better suited to generation one than chasing the latest algorithm
Co-founder Nick met 洪小平 during his first month at DJI. Nick had led large-drone products and early propulsion systems, with strengths in motors, mechanical structures and motor control. 洪小平 focused on sensors, algorithms and overall direction. Together, they broadly covered robotics from perception to actuation.
The first thing they aligned on was not organizational structure but values: “innovate and pursue, embrace growth, build organically, execute well, work with integrity, strive for excellence, stay open and dare to decide.” 洪小平 calls this the company’s “constitution.” It guides hiring and daily work; people who do not fit naturally leave or are asked to leave.
The DJI-style “comfort” he describes is not low pressure. It means a strong technical environment, little office politics and a culture that judges by results. Strong engineers can focus on the problem at hand and challenge experts and established approaches from first principles. Strutt therefore has not lowered its hiring bar at the startup stage.
EV One’s current autonomous-driving stack does not force in the latest embodied model because it cannot operate reliably at low cost, low compute and low data volume. The team uses a technology path closer to early autonomous driving. 洪小平 previously collaborated with Professor Zhang Fu at HKU and has a joint training program. He also cites open-source FAST-LIO becoming SOTA in LiDAR SLAM as evidence that a strong algorithm can remove adoption resistance to new hardware.
22. First-mover advantage only buys time; the real question is how to cross the chasm
洪小平 believes reliability and stability in a safety product are built through brand recognition. The first company to establish a “safe, intelligent” position in consumers’ minds will have a clear advantage. The team has also filed patents on first-of-their-kind issues including joystick control, four-wheel drive and suspension, mode switching and language-model integration.
He does not regard these as permanent moats. A complementary team and complex software-hardware capabilities can slow imitators, but cannot keep the company ahead for 5–10 years on one product. The generation-two and generation-three roadmap must keep “laying eggs along the way” through continued adoption of new technology, or first-mover advantage will be a one-time windfall.
On his tendency to “think everything through as much as possible,” 洪小平 acknowledges the pressure it creates, while also accepting that no architecture can avoid iteration forever. He admires Jobs’s vision and communication on iOS, and 汪滔’s sensitivity to materials, engineering details and product excellence. His goal is to combine top-down strategy with bottom-up innovation.
The biggest question still comes from Crossing the Chasm: who exactly are EV One’s innovators and early adopters, and how does it reach the early majority? One friend has suggested using it as an autonomous mobile camera, an unexpected niche. Drones found the new experience of aerial photography among many small markets. Cheng Manqi warned that a smaller chasm also brings red-ocean competition faster. 洪小平 still prefers a smaller chasm, because that is “difficulty within optimism.”