Vol.185 Industry Watch 34 | Where Are the Non-Consensus Opportunities in AI Companion Robots?
Summary
李丰与何嘉斌把AI陪伴硬件定位为本轮技术周期的“第三波”：2023年看大模型，2024年主要看具身智能机器人（同时也有人讨论Agent），今天则要看谁能交付产品、创造需求并赚到钱。 Once foundational technology has absorbed decades of accumulated “vital energy” and made a major leap, its growth rate tends to flatten; the grandest software-and-hardware stories are also usually the hardest to commercialize, so opportunities spread into vertical applications. “It may not be the most AI-heavy product, but it must involve relatively new technology, scenarios, applications and ways to make money.”
Ropet的非共识不是把大模型塞进毛绒玩具，而是先用可爱、沉默的“弱机器人”获得开机率，再用闭源的行为数据长出智能。 Its camera continuously captures facial expressions, touch, smiles and gestures within roughly half a meter and a field of view of about 120 degrees; if 100,000 users eventually maintain high daily usage, this dataset would exceed the one-on-one pet-interaction samples available from ordinary home surveillance. 何嘉斌’s sequence is explicit: “First you have the hardware, then the behavioral data matched to that hardware, and only then can you call it intelligence.”
产品被定义为“百分百好玩，零分有用”，核心商业选择是避开工具硬件的配置与性价比竞争。 Ropet does not speak human language directly; it uses eye contact, movement and vocal tone to mimic the perception of a pet or a child aged zero to three. Strong conversational ability would make it feel like “a highly intelligent agent stuffed into a plush toy,” or like a two- or three-year-old suddenly arguing with its parents, undermining the sense of being a pet and the trust that comes with it. The team wants to trigger not the dopamine decay that follows novelty, but the urge to care, oxytocin and the satisfaction of “being needed.”
目前接近一万台的交付已经推翻了最初单一的用户画像，显示陪伴产品可能是全球、全年龄市场。 The team initially targeted urban women aged 24 to 30 who lived alone, but more than 50% of its core users in Japan are women aged 50 to 70, while many North American buyers purchase Ropet as a gift for children; roughly half of users still fit the original urban-woman profile. 何嘉斌 therefore calls the company a “future pet company”: desktop, bag-mounted, in-car and roaming home formats could all belong to the same category of “silicon-based pets.”
项目的早期商业验证来自极低预算下两次连续正反馈，而不是先靠融资堆出来的规模。 After four months of keeping the six-person, debt-laden company alive by borrowing from old shareholders, the team spent less than RMB10,000 on a booth at a Shanghai IP licensing fair, where 5 prototypes became the hottest new product. Old shareholders then continued to support the company, while individual shareholders raised roughly RMB10M more. At CES, the founder visited the US for the first time and hired an interpreter for about RMB1,500 a day; the product generated 6B views globally in backend statistics over 6-7 days, making the commercial value of something that could be touched, pre-ordered and delivered visible.
中国消费电子供应链构成了AI硬件最实在的结构优势，但首代产品仍必须承受漫长而昂贵的迭代。 Ropet went through 3-4 major hardware-selection changes from its first prototype to nearly 10,000 delivered units, with each change potentially causing at least a 2-month delay. To create genuine eye contact alone, the team insisted on 2 screens positioned at roughly a 15-degree angle and supporting dual-screen display; it is still working through servo noise, heat dissipation under the plush covering and quiet-fan design. “Be bolder”: building a roughly 50-point framework from mature components first, then iterating quickly through Shenzhen’s supply chain, is what makes mass production at least a year later possible.
智能升级是一条带条件的信任曲线，而不是首代产品把能力拉满。 Today’s 1 TOPS of edge compute must simulate vision, hearing, touch and gravity sensing at the same time. 何嘉斌 imagines that if the device can move to 10 TOPS at the same cost a year from now, while the company accumulates in-the-wild data from 20,000-30,000 units, the pet may grow from “3 years old” to “10 years old” and begin language interaction in an appropriate way. 李翔 contrasts this with the decline in Character.AI engagement and the stepwise progression of autonomous driving from L2 and L3 toward L4 and L5: consumer AI needs to “ship with low expectations first, then climb the AI ladder gradually.”
Deep dive
1. AI Investment Is Moving From Maximum Imagination to a Deliverable Third Wave
何嘉斌 divides technological innovation into 3 stages: foundational technology explodes, capital bets on the largest imaginable opportunity, and startups finally enter specific vertical scenarios, deliver products to users and realize commercial value. By this framework, the theme in 2023 was large models; in 2024 attention shifted mainly to embodied-intelligence robots, while some also discussed Agent. AI hardware represents the search for applications in the third wave.
李丰 adds that every technological leap comes from years of accumulated inputs. This round of large models absorbed 30-40 years of public text data and then added advances in compute and algorithms to produce a “small qualitative change.” Once resources have been concentrated and consumed, the same scale of accumulation cannot be repeated in the short term; foundational progress therefore shifts from explosive to relatively linear.
The most seductive narratives in the second wave were one software, one hardware: Agent could replace every specialist and become a “fully electronic ox” — a digital laborer — while embodied intelligence could replace generalized human labor. But 李丰 warns that the “applications with the greatest imaginable upside” are usually the hardest to land, because they are difficult both to prove and to disprove.
After Tesla’s Optimus plan was delayed as Elon Musk focused on politics, the Spring Festival Gala’s dancing robots, a private-enterprise symposium and the government work report at the Two Sessions pushed China’s embodied-intelligence buzz another notch higher. 李丰’s view is that taking attention to another level is getting harder, which is why specific third-wave applications have become the focus.
2. Ropet Must Become a Pet Before It Can Become an Agent
何嘉斌 defines Ropet as a new form of pet for women, providing emotional value and addressing loneliness. It quietly sits on the desktop and looks at its owner, delivering companionship through continuous “presence” rather than high-frequency conversation. Its facial camera provides an entry point for identifying long-term behavioral and emotional changes.
“The more you raise it, the more it understands you” is not a static personality slogan. The team wants the pet to gradually become a projection of the user’s emotions, learning catchphrases and behavioral habits; if cared for properly, it might even start calling the user “Mom.” 何嘉斌 believes what high-pressure urban life truly lacks is “that small feeling of love.”
The company’s long-term positioning is “a future pet company,” not merely a maker of a single desktop toy. 何嘉斌 imagines that future pets will include a “silicon-based pet” species called Robo Pet, which could sit on a desk, hang from a bag, enter a car or move around the home.
3. “Make People Like It First” Was the Key to Changing 李丰’s Mind
李丰 had previously reviewed 3-5 projects in a row and came away with the same impression: “making a toy for AI.” The technology was highly visible, but the products did not look as if they started from consumer demand, so he had decided to stop reviewing AI toys. Several colleagues insisted that he meet 何嘉斌, which changed his view.
The real difference was not how much AI the product used, but that 何嘉斌 first considered “how to make people like it” and only then decided how much functionality AI should carry. He clearly established the product as primary and technology as secondary. 李丰 believes this is more likely to define a product consumers are willing to keep than starting with an AI capability demo and looking for a toy shell afterward.
李丰’s initial question was how a male founder could think so concretely about women’s emotions and product details. The answer was not an innate ability to “understand women,” but 何嘉斌’s 10 years working at the intersection of women’s consumption, aesthetics, human-computer interaction and consumer hardware.
4. AI Glasses Show That Global Hardware Cannot Be Detached From Local Lifestyles
李丰 uses Meta’s partnership with Ray-Ban on AI sunglasses as an example. More than 80% of American adults drive themselves to work, with one-way commutes lasting over 40 minutes and close to 50 minutes; users might already wear sunglasses for roughly 2 hours a day. Taking them off to charge when the battery runs low does not conflict with existing habits, making sunglasses a “very clever” form factor.
Chinese consumers face different constraints. Myopia is widespread, and prescription-glasses users must deal with vision correction, weight, exercise and all-day wear. Adding a battery, camera, chip and sensors — then requiring users to remove the glasses every 2 hours to charge — imposes a much higher cost of lifestyle change.
李丰 cites data from a portfolio contact-lens company: in 2020, only about 7% of buyers were not nearsighted and wore contacts purely for appearance; over 90% of users were women and nearly 10% were men. By 2024, the company’s sales had reached more than RMB1B, and the non-nearsighted share had approached 30%. Asian women may dislike frames but also may not want AI glasses to cover the colored contacts they bought for aesthetic reasons. 李丰 therefore has not made AI glasses a major focus, while stressing that this is only “one company’s view, a partial perspective.”
5. Nearly 10,000 Deliveries Overturned the Original Narrow User Profile
The team initially defined its users as urban women aged 24 to 30 who lived alone. They were heavy opinion leaders on Xiaohongshu and Instagram and willing to share new lifestyles. But after nearly 10,000 devices were delivered over time, the real demand proved more dispersed than the profile suggested.
More than 50% of core users in Japan are women aged 50 to 70; in North America, many people buy Ropet as a gift for children. Roughly 50% of users still match the original urban-woman definition, but 何嘉斌 realized that “pets do not belong to any particular age group.” Companion products may naturally address a global, all-ages market.
The desktop remains an important first-generation scenario. From graduating college to leaving the workforce, a person may spend 30-40 years with much of their time at a desk. It offers stable power, a computer and a phone, and is the easiest place to maintain presence and daily usage. A “desk companion” is therefore not a decorative concept but a usage-frequency design.
何嘉斌 acknowledges that hardware cannot leap over delivery cycles, and the first generation can rarely complete the full set of insights. Usually the first batch must ship before 2-3 further generations of corrections can bring the product close to a breakout hit. He compares it with the iterations from the 3GS to the iPhone 4: new categories usually explode after, rather than with, the first product.
6. A Decade of Work Built the Capabilities Behind “The Beauty of Uselessness”
何嘉斌 graduated from Beijing Institute of Fashion Technology in 2014. The school’s programs in fashion, jewelry, apparel and bags all centered on understanding female consumers. As an industrial-design student, he chose to build a cross-disciplinary advantage through technology, believing that “technology will definitely be the primary productive force,” but that technology must be wrapped in a layer of aesthetics.
He then joined Microsoft Research Asia to study human-computer interaction. A decade ago, he was already exploring the first-person assistant now called AI glasses; one question was how to make women willing to wear a camera on their heads. He later entered Baidu’s deep-learning lab, spending 2 years exploring lab-grade innovative hardware.
In his first startup, he co-founded Ling with his former leader and spent 5 years making children’s companion robots, selling “several million units” in total but also encountering the growth ceiling of the category. From 2021 to 2023, he joined ByteDance to fill in his experience with large projects and supply chains, participating in Pico 4 and experiencing what it meant to “spend tens of billions of RMB making one product.”
The period when ChatGPT moved from 3.0 to 3.5 triggered his return to entrepreneurship. 何嘉斌 judged that AI opportunities would land in hardware carriers capable of collecting data and receiving behavioral feedback, not necessarily in bipedal humanoids. He chose “the beauty of uselessness,” the area best suited to absorb his decade of accumulated experience, combined it with 真值创投’s technology in biomimetic machines and assembled a team in roughly 3 months.
7. Technology Cycles Always Fall Back From Revolutionary Narratives to Stepwise Deployment
何嘉斌 believes that once foundational technology matures, startups must answer the question: “What do you apply it to?” Agent can tell a story about replacing intelligence; robots can tell a story about entering homes and factories. But the further the cycle progresses, the more it must solve concrete problems such as activation rate, time spent, delivery and commercial value.
李丰 uses autonomous driving to review the same pattern. Before 2015, Google and Waymo, before its split, were considered leaders because of the scale of their road testing. After 2015, Uber had a global dispatch network and Tesla had large numbers of mass-produced cars fitted with sensors, so they became the vehicles for the grandest commercial visions.
Some even promoted the idea that Tesla cars equipped with autonomous driving could run ride-hailing trips for owners while idle; if Uber achieved driverless operation, it would no longer need to share revenue with drivers. Once the narrative returned to reality, the standards became who had actually put the technology into mass-produced cars and which automakers they were working with. China, too, is still moving from L2 toward L3.
About 11 years after the first autonomous-driving bubble in 2014, the industry remains at the L2-to-L3 stage. Higher-level systems instead appeared first in closed, driverless environments such as campus delivery, ports and mines. 李丰’s point is that the third wave does not mean grand visions disappear; it means those visions are broken down into verifiable small scenarios.
8. The Biggest Asset of New Hardware Is Behavioral Data That Did Not Exist Before
何嘉斌’s route is not to stuff existing AI into hardware and then look for a scenario, but to build the scenario and product first and prepare data for future AI. Conversation data can be captured through the microphones on phones and computers; there is no existing dataset for the long-term behavioral interaction between a pet and its owner.
Ropet’s camera is aimed at a distance of roughly half a meter and a field of view of about 120 degrees. The goal is to capture one person’s facial expressions, gestures, touch, smiles and even microexpressions. Home surveillance also produces visual data, but it usually covers a broad scene, does not focus on a single person and does not capture a long-term interaction relationship at such close range.
何嘉斌 offers a conditional projection: if the company eventually has 100,000 users and devices maintain high daily usage, with users continuously touching, smiling at and caring for their pets, these closed samples could be used to train a model of the pet’s stress responses. “First you have the hardware, then the behavioral data, and only in the third step do you have intelligence.”
9. Weak Robots Use “Being Cared For” Instead of Showing Off Strong Intelligence
The team deliberately did not connect Ropet to an end-to-end conversational large model. Instead, the pet expresses emotions through movement and vocal tone. 何嘉斌 believes that once a pet speaks fluent human language and displays logical reasoning, preferences and linguistic boundaries, users may see it as an agent stuffed into a plush toy and lose their sense of trust in it as a pet.
李丰’s analogy is that if your own two- or three-year-old child suddenly started arguing with you, your perspective on and feelings toward the child would change dramatically, even though the child was still your own. Capability is not always better when it is stronger; a mismatch between capability and role expectations can itself create unease.
何嘉斌 borrows from Japan’s concept of the “weak robot,” positioning the product as a child aged zero to three or a pet, rather than a “chatterbox” in the home. He says Japan is at least 20-30 years ahead of China in robotics, both academically and industrially, and its experience shows that highly intelligent devices are actually harder to integrate into daily life over the long term.
This choice corresponds to 2 entirely different retention mechanisms. Products that stimulate dopamine fade after the novelty period and eventually gather dust; pets that continuously awaken the urge to care may trigger oxytocin, create the satisfaction of “being needed” and maintain daily usage. “Making it weak” is not about doing fewer things; it is a design for stickiness.
10. Smartphones Created New Data First; Platforms Created New Business Models Later
李丰 asks why mobile-internet giants often succeed only on their founders’ third startup. It was not that they had previously been unwilling to succeed; they were waiting for smartphones to become widespread. Before new data such as location, images and high-quality audio existed, ride-hailing, food delivery, photo editing and short video lacked their most important means of production.
Compared with Nokia and Motorola, smartphones added high-definition optical cameras, MEMS microphone arrays, gyroscopes and GPS. Platforms could therefore simultaneously know the locations of consumers, delivery riders, restaurants and vehicles, while also accessing images and sounds that could be distributed. Only then did “the relationship between 2 locations” become ride-hailing or food delivery.
The sensors were not installed by platform companies, but platforms waited until the hardware became widespread before building business models on the new data. 李丰 believes AI companion hardware is repeating the same relationship: the task today is to put sensors into new life scenarios; tomorrow, products and consumption models unsupported by phone data alone may emerge.
11. China’s Supply Chain Makes “Application-Driven Technology” an Executable Advantage
李丰’s view of China’s AI hardware advantage rests on 2 industrial chains: one for sensors and chips, the other for precision processing and consumer-electronics manufacturing. Add AI software, and China may form a software-and-hardware combination that other countries cannot easily replicate at the same time. He calls this “application-driven technological innovation.”
Robot vacuums have already reached a state in overseas markets where “Chinese companies compete against Chinese companies.” 李丰 thinks new-energy vehicles may follow the same path; if Huawei phones had not been restricted in 2019, smartphones might also have moved in that direction. The common factor is not leadership in a single algorithm, but the ability of software to be embedded quickly into a scaled hardware supply chain.
Insta360 is a case 李丰 cites repeatedly. After struggling to sell software in China, the company shut its Nanjing office and moved to Shenzhen, embedding its algorithms into 360-degree cameras. The product entered Apple’s physical stores in just over a year and later grew into what 李丰 calls a company worth “more than RMB100B today.”
An embodied-intelligence scientist teaching in the US once told 李丰 that when a US team iterates on a robot, any hardware change in a single technical component requires a new iteration, so it can only go to China for custom manufacturing; each round takes roughly 3-4 months. After returning to Shenzhen, a requirement sent out in the afternoon might come back the next evening. 李丰 acknowledges that the “10 years versus 1 year” comparison is exaggerated, but the iteration-speed gap is real.
12. The Company Had to Abandon Expensive Complex Robots Before Finding an Opportunity Below RMB2,000
When 何嘉斌 left ByteDance, he was product VP at Mengyou. There was initially another CEO, and the team wanted to use its supply-chain capabilities to build a complex robot modeled on Japanese products that could move around the home and perform multiple tasks. At the time, he preferred to focus on the product work he knew best and was not prepared to become the top executive directly.
Long-term user testing delivered a brutal conclusion: for the small group willing to adopt early, there was little difference between a product priced at RMB10,000, RMB70,000 or RMB100,000. To become a breakout hit, however, the price had to fall below roughly RMB2,000. The company therefore shifted from a technology-and-supply-chain orientation to one centered on product and user experience.
The original CEO favored the technology and supply-chain route and eventually left; 何嘉斌 was “called in at a moment of crisis” to take over as CEO. Embodied intelligence was still a capital-market hot spot, making this company — which was deliberately narrowing its capabilities and pursuing emotional value — look “highly non-consensus.” Fundraising was not easy.
何嘉斌 took on company debt and borrowed from old shareholders, extending the new strategy for 4 months. The product direction was not chosen leisurely with ample cash; it was a “do-or-die battle” completed while both cash and the organization were under pressure.
13. A Booth Costing Less Than RMB10,000 and 5 Prototypes Saved the Company
Last October, the 6-person team brought 5 prototypes to a Shanghai IP licensing fair. The entire booth, including venue rental and construction, cost less than RMB10,000. Ropet became the fair’s hottest booth in the 4 years since the pandemic; almost everyone passing by stopped to ask, “What is this?”
何嘉斌 attributes the impact to “making AI visible” and a “visual hammer.” Eye tracking, feedback after being touched and pet-like movements made Ropet visibly different from ordinary plush toys. Visitors could feel that it was an AI product without necessarily being able to explain where the AI was.
The validation brought the first key funding after 何嘉斌 became CEO. Old shareholders continued to support the company, while individual shareholders he met during a Shanghai roadshow contributed roughly RMB10M more. The team could finally open molds, find a factory and push toward mass production, while also having enough budget to attend CES for the first time.
CES delivered a second round of positive feedback. 何嘉斌 was visiting the US for the first time and his English was “particularly poor,” so he hired an interpreter for roughly RMB1,500 a day for 5 days. Traffic surged after Forbes conducted an interview on the first day; users at the booth could touch, hug and interact with Ropet, then pre-order it through KISS. Backend statistics showed 6B views over 6-7 days, also leading to follow-on fundraising contacts including Fenghe Fund.
14. First-Generation Innovative Hardware Can Only Start From a “50-Point Framework”
何嘉斌’s most direct lesson is that “it is hard to think everything through in the first generation.” The team first selected mature components it could manage and source within controllable lead times, built a roughly 50-point hardware framework and moved forward quickly, rather than waiting for a theoretically perfect but industrially uncertain solution.
From the first prototype last June to nearly 10,000 delivered units, the product went through at least 3-4 major hardware-selection changes. The platform might remain the same, but compute, cost and peripheral configurations continued to change. Each platform change could create at least a 2-month delay and overturn a batch of designs already completed.
The eyes show how details can dictate system selection. A flat 4-inch screen with a cover looked “particularly like a robot”; 2 screens set at roughly a 15-degree fold created a more three-dimensional sense of spatial eye contact. The chip therefore had to support independent dual-screen display, while the early platform could drive only a single screen.
Servo noise, fan selection and heat dissipation under the plush cover also required repeated iteration. A fan costing a few RMB might be replaced by a more expensive quiet solution. 何嘉斌 estimates that moving from a demo to mass-producible delivery takes at least a year; 李丰 believes the mature Pearl River Delta supply chain is what makes that speed possible.
15. “100% Fun” Means Deliberately Giving Up Tool Value
Asked to score the product on usefulness versus fun, 何嘉斌 gave an extreme answer: “100% fun, 0% useful.” Usefulness here means efficiency and tool value; fun means providing pure emotional value. The 2 can coexist, but they dilute each other when built into the same definition.
His earlier Luca picture-book-reading robot was a 50-50 product, or even 30-70: parents bought it to satisfy the practical need of building reading habits in children aged 3 to 6, while children used it for the IP, interaction and game elements. Because the buyer and user were separate, the product had to please 2 groups at once, making its definition prone to “arguments.”
何嘉斌 cites 王宁’s counter-question: “If LABUBU were a USB drive, would you buy many USB drives?” Once users hold even “the slightest expectation” of tool value, fun and aesthetics are discounted. Emotional value now has stronger market power and pricing power, making pure “uselessness” easier for the first time to turn into a business model for technology consumer goods.
Tamagotchi proved that a nurturing game can become globally popular on its own: players feed the pet, clean up its waste, prevent it from getting sick or dying, and repeatedly receive feedback from care and growth. It can still attract users after being re-collaborated with Miniso. Ropet aims to bring the same sticky core into the physical world, using AI and five-sense interaction to become more biomimetic rather than adding efficiency features.
16. Low-Cost Models Let a 6-Person Team Assemble “4 Senses” on 1 TOPS
何嘉斌 calls the fundamental change from 5-6 years ago the “cost of entrepreneurship.” In the past, even recognizing a fruit might require a team to collect data and train a model itself; today, a startup can obtain models from the open-source GitHub community, then quantize, prune and deploy them on the edge.
Ropet’s current edge chip has only 1 TOPS of compute, yet must simultaneously simulate vision, hearing, touch and gravity sensing, while also dealing with heat dissipation blocked by the plush cover. A 6-person team can assemble these models in 4-5 months because the cost of obtaining models has fallen.
李翔 specifically attributes the increase in open-source supply to DeepSeek breaking the closed-source monopoly, making open source a commercial choice more manufacturers must adopt. For small teams, the value is not merely access to a stronger large model, but the fact that large numbers of vertical models can finally be selected, pruned and combined.
17. A Pet Can Grow From 3 to 10, but It Cannot Skip the Trust Ladder
何嘉斌 estimates that current hardware and models can simulate at most the perception of a child aged zero to three: recognizing a limited range of objects, using faces to judge basic emotions and responding to gestures and touch. When a user makes a “shh” gesture, it becomes quiet; when its head is stroked, it actively resumes the interaction and turns its voice back on.
Higher intelligence is a conditional scenario. If edge compute can rise from 1 TOPS to 10 TOPS at no additional cost a year from now, while the team accumulates in-the-wild data from 20,000-30,000 devices, the pet might grow from “3 years old” to “5 years old” and even “10 years old,” then begin language interaction in a way consistent with its role.
The team is not opposed in principle to a pet speaking. It believes that directly connecting an end-to-end large model today would mismatch the interaction model and make users “break character”; 何嘉斌 does not even rule out adding language next year. The standard for releasing capability is not whether the model can do something, but whether the product has built enough data, trust and continuity of role.
李翔 cites declining engagement over the past 6 months at Character.AI and similar companion apps as a counterexample. If open-ended language generation were enough to solve broad emotional companionship, a phone and OpenAI would already be sufficient. Just as consumers must move step by step from internal-combustion cars to EVs, automated parking and hands on wheel before reaching hands off wheel, and regulators must move from L2 and L3 before discussing L4 and L5, AI companionship also needs to “ship with low expectations first, then climb the AI ladder gradually.”