I Found the First Companion Robot I Actually Want to Buy! | A Conversation with 世博, Founder of 越伴动力 [Road Podcast]
Summary
- 月半动力 has launched 小伴, a consumer-grade bipedal companion robot standing 65 cm tall, weighing 12 kg, and occupying just 0.12 square meters (smaller than a robot vacuum), priced in China at RMB8,499 with a limited-time presale price of RMB6,499 and delivery as early as the National Day holiday. 世博, the founder, gives the segment’s cool reception a one-line explanation: “No genuinely good product has appeared” (没有真正好的产品出现). Pricing it below RMB10,000 is meant to let consumer feedback show the market what a companion robot with real vitality should look like.
- The product philosophy is “less is more” (少就是多): across 300+ surveys, 90%+ of users wanted the robot to speak, but segmentation showed they were not asking for a “cold, AI-like encyclopedia”; 小伴 therefore does not speak human language, instead using a self-developed 12-sound language based on Swahili’s CV structure, with body language carrying the main emotional load (the Mehrabian experiment puts the split at 55% body language, 37% tone, and only 8% content). 世博 has kept his Lovot powered on since he bought it: “If it could say even 1 word or 1 sentence, I might not have bought it.”
- The stack uses a 3-layer VRA framework (the industry generally uses 2): an edge-side fast brain with 1.7B parameters handles instantaneous decisions, a 7B slow brain handles complex reasoning, both fine-tuned from an open-source Qwen model with interaction latency kept under 0.4 seconds; a cloud-side 200B+ model and ultra-long-term memory drive multidimensional evolution of personality parameters. The industry has no real in-home data, so the team built a “family simulator”—essentially a world model—to generate ultra-long-horizon emotional-interaction data for fine-tuning: a synthetic-data-first training strategy.
- The bipedal form is a first-principles choice: “20 years from now, every household will have one, and it will definitely be bipedal”—only bipedal robots can navigate stairs and other unstructured terrain, and their movement expressiveness is “not even in the same order of magnitude” as wheeled or tracked robots. Safety engineering includes a silicone-like soft shell covering 95% of the body, thin-film pressure sensors in the legs (“by then, current-based motor torque feedback would already be too late”), and a deliberately compact form: a robot weighing tens of kilograms falling onto a child or kitten would be “very dangerous.”
- His technology-roadmap verdict is blunt: the “VLA is dead” argument is “exactly right,” and its ceiling is broadly predictable; world models are “a reasonably credible route for the industry right now,” but he questions whether turning implicit prediction into explicit training creates a more fundamental difference and says whether it can be operationalized is “hard to say.” One emergent result after fine-tuning: when asked “Do you want to go out and play?”, 小伴 could answer only “want to” or “don’t want to” before fine-tuning, but answered “Your call” (看你) afterward.
- The retention design is entirely about “vitality,” not functionality: it rejects “party-trick interaction” (call it ugly and it gets teary-eyed, looks wounded, and runs away instead of lecturing); its personality co-evolves with its owner’s E/I profile and it tags people individually; it deliberately offers no way to view or edit its parameters and uses no removable battery—“It is an individual … this really breaks the spell.” The 3 target user profiles are middle-aged women drawn to the experience of raising a companion, families buying it for their children, and tech early adopters.
- The execution signal is clear: 15 full-time employees and 10 interns, a launch pace faster than 90% of robotics companies, a 真格 seed round raised late last year, and a 26-year-old founder who has worked on nearly 30 robots and previously served as product partner at 高擎动力. His self-assessment: “Extreme passion, extreme efficiency, plus the fact that we had almost never made a mistake before.” If even 1 technology-route choice had been wrong, there would be no fully interactive companion like this today.
Deep dive
1. 26 Years Old, Nearly 30 Robots, and a Launch Pace Faster Than 90% of Peers
世博 took a leave of absence from his PhD at Beijing Institute of Technology, is an INTP Capricorn, and has been building robots for 8 years. “I’ve explored all the classic configurations—wheeled, legged, tracked, and robotic arms,” including an adult humanoid measuring 1.7 m (1 m 7 dm) and weighing 70 kg, plus a heavy-duty quadruped. He started selling quadruped education kits in his 2nd year of college and says, “By unit sales, I should have sold more than 王兴兴.”
He previously served as a product partner at 高擎动力, leading a 30-person product and R&D team spanning mechanics, hardware, embedded systems, motion control, and perception. After raising a 真格 seed round late last year, the team now has 25 people and “has already launched its first companion robot, probably faster than 90% of robotics companies.”
He attributes the pace to a combination worth noting: “Extreme passion, extreme efficiency, plus the fact that we had almost never made a mistake before, including in our technology-route choices—if we had chosen even 1 path incorrectly, you would not be experiencing an interaction this complete.”
2. Why Companionship: He Couldn’t Be Top-Tier in Manipulation, but Could in Companionship
Starting in 2022, he explored companion robots on his own at home, first building a 2-wheeled robot—“1 cable could get it stuck”—before settling on a bipedal design in 2024. His candid rationale for choosing the field: manipulation robots were the hottest segment. “I could do it, of course, but I definitely couldn’t be top-tier; in emotional companionship—non-manipulation, fully soft-bodied bipedal robots—I can be top-tier,” he says, attributing that edge to “extreme passion, extreme love.”
The underlying motivation is personal: he keeps a cat and “really wanted to build a robot that could keep me company.” His cat cannot fully understand his emotions, “but a robot is different—it understands what I am saying and gives better feedback on my emotions and behavior.”
3. The Lesson from Lovot: Less Is More
Before founding the company, 世博 bought a Lovot that has stayed powered on ever since. He is struck by its minimalist interaction: when it spots its owner, it rolls to his feet, retracts its wheels, and opens its little arms to ask for a hug—“just 1 simple interaction, exactly enough.” His key line: “If it could say even 1 word or 1 sentence, I might not have bought it.”
His course correction is candid: he once thought older companion robots—he bought both Vector and Emo—were “too poorly made” and entered the field intent on making something smarter, with more features. Now he says, “when it comes to companionship, less is more; a good product is not built by piling up features.”
More than 300 surveys and dozens of interviews found that over 90% of users wanted the robot to speak, but segmentation showed they wanted not “a cold, AI-heavy, question-and-answer encyclopedia” but “a genuinely living companion”—hence neither the industry’s cookie-cutter announcer voice nor unintelligible onomatopoeia.
4. A 12-Sound Alien Language, with Body Language Carrying 55%
小伴’s language consists of 12 sounds and borrows the CV structure of Swahili. The team deliberately looked to an African tribal language because it plans to go overseas: “It can’t sound too much like Chinese or too much like English.” The expressions are intent-based: when it needs to recharge, it pats its belly and says “I want to eat” (“迷它哭”); “他o他推” means “let’s go outside and take a look.” Spend an afternoon with it and “you gradually begin to realize what it is saying.”
The design follows the Mehrabian experiment: body language accounts for 55% of interaction, tone 37%, and spoken content only 8%, so the body is the primary channel for emotional expression. The arms are deliberately short and use a relatively unique continuous-deformation design instead of joints; “we don’t want it to perform manipulation tasks,” only express emotion.
The eyes use 6-layer skeletal animation: the pupil, sclera, and pupil pattern are parameterized independently to create “the breathing quality of a naturally flowing gaze.” Like human pupils, they tremble and flicker; when it looks at something, its gaze does not move at a constant speed.
5. An Alien Cutie: Not a Cat, Dog, or Human, and Genderless
The positioning logic is simple: “The more you make it look like a cat or dog, the easier it is for real cats and dogs to beat it.” It is not humanoid either: “we don’t want to set user expectations too high, whether on intelligence or performance.” Its body is 95% covered by a silicone-like soft shell, with a belly up to 2 cm thick—something “many companies in the industry want to do but have never actually pulled off.”
It has no gender; users define it through accessories: magnetic ears, bow ties, and newly made 梅西 and C罗 jerseys. It can also blend its owner’s voice with an idol’s into a “genetic timbre”—“like the child of you and your idol.”
6. Against “Party-Trick” Interaction: How to Build Vitality
A party-trick interaction is “小A小A, do a dance, shake hands.” In a competitor test, calling the robot ugly drew “It is wrong to speak to people that way,” followed by a lecture; 小伴 instead gets teary-eyed, lowers its head, hugs its belly, and even stomps, turns away, and runs off—“you have to be able to feel this for it to feel like real companionship.”
Its personality develops across multiple dimensions: “the purpose of cloud-side ultra-long-term memory is to drive multidimensional growth and dynamic evolution of personality parameters.” An E-type owner raises a more E-type 小伴; if an I-type owner consistently does not respond, it will “become less willing to seek interaction with you and pay more attention to household objects, flowers, and plants.” Violent children are tagged and gradually receive less proactive attention, “but what will not change is this”: when a child comes to it, it will not immediately run away.
On day 1 out of the box, it is “extremely curious”: it walks around the entire home to build a map and tries to squeeze through a half-open bedroom door. The initial getting-to-know-you period lasts about 1 week; afterward it wanders back and forth, greets selectively, and is willing to come very close.
7. Fast Brain, Slow Brain, 0.4 Seconds, and the Family Simulator
The brain uses a 3-layer VRA framework, versus the industry’s usual 2 layers: the edge-side fast brain has 1.7B parameters and handles instantaneous behavior decisions; the slow brain has 7B parameters and handles complex logical reasoning, with latency already controlled within 0.4 seconds. The base model is open-source Qwen, including a cloud-side 200B+ model; Qwen was chosen because “at a very small parameter count, its reasoning and understanding of human emotion perform relatively well.”
No robotics company in the industry has real in-home deployments or real interaction data, so the team built a “family simulator” that models 小伴’s ultra-long-horizon active and passive emotional interactions with environmental objects, family members, and even cats and dogs after it enters a home; after labeling, the data are used to fine-tune the edge model. Emotion recognition relies on context rather than a single frame: “what you have done over the past 1 hour is combined to infer your current emotional state.”
For privacy, the edge collects no raw audio or images; it trains only on semantic embedding features. On the question of timing, the odds of building 小伴 3 years ago were “fairly low,” given the advances over the past few years in language models, vision models, VRA, world models, and related architectures.
8. Bipedal Is a First Principle; 65 Centimeters Was Calculated
“20 years from now, every household will have 1, and it will definitely be bipedal.” From first principles, only bipedal robots can navigate unstructured terrain such as stairs, and their movement expressiveness is “not even in the same order of magnitude” as wheeled or tracked robots. Something meant to live in a home long term and build an emotional connection “should rely on feet and legs.”
The 65 cm/12 kg specification was calculated, not guessed: its maximum outdoor running speed is 1.9+ m/s (adults walk at 1.1-1.3 m/s; “as long as you don’t run, it can keep up”); it can climb steps up to 15 cm—the national standard for public steps caps them at 15 cm, and most are around 10 cm. Making it taller and larger would be less safe: “When something weighing tens of kilograms falls and hits your cat or child, that is a very dangerous situation.”
Safety redundancy includes thin-film pressure sensors in the legs: if a child is touching the robot when it suddenly stands up, relying only on current-based torque feedback from the joint motors would mean “it is already too late.” He also argues that safety and privacy regulations should have been established early in the industry’s development: “Robots are not like cars; cars have clear, single-purpose functions.”
9. Companionship Is Not the Same as Pleasing
世博’s framework is that “pleasing is a relationship that is neither sufficient nor necessary for companionship.” The real focus is “proactive interactive empathy”: understanding your current state and delivering the appropriate interaction and feedback. When called ugly, 小伴 expresses only its own hurt, turns away, and runs off; refusing to please is precisely where the feeling of companionship comes from.
Interaction does not use a fixed action library: even in 2 identical scenes with exactly the same words, it will not give the same response. It generates 5-10-second stylized motion sequences based on context and emotion—a goodbye wave can be excited or bashful.
10. It Is an Individual: No Removable Battery, No Parameter Access
Battery life is 3-4 hours indoors and just over 1 hour of continuous walking outdoors, but the battery is explicitly non-removable: “It is an individual; you should not take part of its body out and put it back in—this really breaks the spell.” By the same logic, users cannot view or edit its personality prompts or tags—technically “very easy to implement,” but the team wants users to treat it as an independent, living individual, a member of the family.
The charging scene is also made lifelike: a cute little seat lets its small hands clutch its belly as it lowers its head, as if resting. The host offered a parallel: a Japanese plush robot, apparently Moflin, sells for RMB4,000; to preserve the feeling of a living creature, it has no plug at all and can only charge wirelessly through its base.
11. Pricing, 3 User Types, and “2 Base Capabilities”
The China price is RMB8,499, with a limited-time, limited-quantity presale at RMB6,499 and delivery as early as the National Day holiday. Keeping it below RMB10,000 is meant to “reach more ordinary users and, through consumer feedback, tell the market what a good, vital companion robot should look like.” The segment had not been viewed favorably; “the reason is clear: no genuinely good product had appeared.”
There are 3 target user groups: unmarried middle-aged women who want to experience “raising a child” (小伴 is a raise-and-grow companion with personality evolution and long-term memory), families buying it for their children, and tech early adopters. It has no strong functional attributes and is aimed solely at emotional well-being and companionship; the company is building 2 foundational capabilities—a household-robotics base and an emotional-interaction-model base—which together “are what we call a robot’s vitality.”
A Lovot user event in Shanghai showed him that its users are “a refined, predominantly female group”: they dress their robots and talk about surprising moments at home. Raising a robot instead of a cat or dog “is a lifestyle choice.” His kitten was recently diagnosed with polycystic kidney disease: “I was devastated; it took me a long time to recover.” The pain of saying goodbye to a pet “may be 1 reason people want companion robots.” The next step is to bring the robot to 10 kg or less.
12. “VLA Is Dead”? A View from Someone Shipping
小伴’s emotional brain is a VRA for high-level decision-making, while the manipulation VRA for low-level manipulation is downstream: “1 decides what to do, the other decides how to do it.” 小伴 does not need to perform tasks today, so it solves “what to do” first; separate action-expert and language-expert models support it.
On the “VLA is dead” claim: “I think it is exactly right; you can basically predict its ceiling.” World models are “a reasonably credible technical route in the industry right now,” but he is skeptical that turning implicit prediction into explicit training creates a more fundamental difference in technical principle, and says whether it can be operationalized is “hard to say.”
The family simulator is itself a world model. One emergent fine-tuning result: when asked 小伴, “Do you want to go out and play?”, before fine-tuning it would answer only “want to” or “don’t want to” no matter how it was asked; afterward it answered “Your call” (看你; in 小伴’s language, “O V”). It had learned to infer the questioner’s underlying intent.
13. Loaches, Tears, and Doraemon
A personal footnote: his signature “No Excuse for Engineer” came from an engineering-drawing teacher in his 1st year: “When you run into a problem, don’t make excuses; analyze the cause and make sure it does not happen next time.” The line has stayed with him for 8 years—“this is not harshness; it is an attitude toward building robots.” As a child, he kept loaches for several years, but poured them into a ditch to make room in the tank for goldfish bought from a street vendor: “The moment it slid out of the fish tank, I deeply regretted it, and I still do.”
His vision in 1 line: “Who wouldn’t want Doraemon to step into the real world? Who says only 大雄 of childhood can have Doraemon?” The most meaningful moment came during filming, when 2 children, without any instruction, chased and played with the robot for nearly 1 hour until it ran out of power. When the robot first stood before him in its complete form, “the tears came out on their own”—he felt moved, surprised, satisfied, and proud, “and of course I couldn’t let the team see it.”
On the industry map, Bilibili followers call him “a young version of 智慧君.” His pointed response: “We each focus on 1 task; with 1 person building 1 robot, who knows who will do it faster or better?” But the two are close: last Sunday, 智慧君 had someone call to praise 小伴 and arrange an in-person meeting in Shanghai. The founder he admires most is Boston Dynamics founder 马克·雷伯特: “He started building autonomous robots 20 years ago and is a pioneer in the industry.”