AI Companion Robots & Robot Pets: The Ultimate Guide

Consumer companion robotics has a worse survival record than almost any other hardware category, and it is simultaneously the fastest-growing consumer AI hardware segment in China. Both statements are true at once, and the gap between them is not funding or technology. It is a small number of design decisions that are visible from outside, before you buy.

What follows: what killed the previous generation, what the current wave does differently, which company is behind each brand and where it actually sits, and the two engineering approaches to putting a face on one of these machines.

Companion reading: how to choose a humanoid robot, tactile sensing and robot e-skin.

Table of contents

1. Key takeaways
2. The graveyard, and what it costs
3. The five ways they die
4. The phone test
5. Who is shipping in 2026
6. Where these companies actually are
7. Two ways to build a face
8. What a face costs to manufacture
9. How to evaluate one as a buyer
10. Frequently asked questions
11. Sources
12. Changelog

The graveyard, and what it costs

The category's failures are unusually well documented because they were unusually well funded.

Jibo raised roughly $73m, shipped late into a market Amazon had already commoditised with the Echo, and in 2019 sent its robots a farewell message before the servers went dark. The hardware kept powering on. The character did not.

Anki raised $205.26m across 11 rounds and built the best-executed consumer robot of its generation in Cozmo, which was the top-selling toy on Amazon in 2017. It closed in April 2019, citing a lack of significant funding to support both a hardware and a software business. That is the detail worth sitting with: commercial traction did not save it, because the unit economics and the capital structure were the problem.

Kuri was cancelled by Mayfield Robotics before meaningful volume. Moxie, the $800 child-focused companion from Embodied, announced closure on 26 November 2024 after a lead investor pulled out of a funding round, and switched its servers off on 30 January 2025. Because the conversation ran entirely in the cloud, the robots became inert on that date. There were no refunds except for purchases made in the previous 30 days. A volunteer project, OpenMoxie, later brought some back on local servers, which is the clearest possible demonstration of the argument in this guide. Humane sold its assets to HP after the AI Pin failed, which is not a companion robot but shares the same failure signature: a device dependent on a service that outlived neither.

The five ways they die

1. The cloud switch-off. The personality is a server. When the company dies, the product does.
2. The phone already does it. If the value is conversation, a phone delivers it with better microphones, a better screen and no additional purchase.
3. The novelty cliff. Engagement that peaks in week one and collapses by week six. This is the failure mode Ropet's claimed 80 percent 90-day retention is aimed squarely at.
4. The general-purpose trap. A robot that promises companionship, security monitoring, education and smart-home control does all four badly and costs enough that any single failure feels like fraud.
5. Unit economics that never close. Hardware sold near cost with an assumed subscription attach rate that never materialises. This is what quietly killed several of the above.

The phone test

Before buying, or building, apply one question: what does this do that a phone running the same model cannot?

There are only a few honest answers. It occupies physical space and so has presence a screen lacks. It has embodied motion that reads as intention. It is always-on in a room rather than in a pocket. It is safe to hand to a child or an elderly relative in a way an unlocked phone is not.

If a product's pitch does not rest on one of those, the hardware is decoration on a chatbot.

Who is shipping in 2026

Three of those deserve a second look, because each solves the novelty cliff by refusing to be the object of attachment.

BubblePal is not a toy. It is a pendant that clips onto a plush toy the child already owns, which means the attachment predates the hardware and survives it. Haivivi reports more than 250,000 units sold and a RMB200m Series A, which makes it one of the category's genuine commercial successes.

LOOI is a dock. Your phone supplies the face, the compute and the connectivity, and the dock supplies motion and presence. It sells for about $150, entered mass production in November 2024, and has shipped more than 10,000 units with about 90 percent going overseas. It is also the only product here that structurally passes the phone test, because it does not pretend the phone is not involved.

Fuzozo put a cellular radio in a bag charm, which takes the companion out of the house and away from home Wi-Fi. It took more than a thousand JD.com pre-orders in ten minutes and appeared at CES 2026 with Tuya supplying the cloud layer.

For scale on the market itself: roughly RMB20bn of capital has moved into Chinese AI toys, Shenzhen AI toy sales grew about 395 percent year on year, and Ropet reports 90-day user retention above 80 percent, which if it holds is the single most important number anyone in this category has published, because retention is exactly what the previous generation could not achieve.

The pattern in the successful end of that table is restraint. Moflin has no face and cannot speak, which removes the uncanny-valley problem and most of the cloud dependency at once; Casio prices it near $430. Its lineage runs back to Yukai's Qoobo, a cushion with a tail and no face at all. Moflin has no face and cannot speak, which removes the entire uncanny-valley problem and most of the cloud dependency at once. Loona commits to being a pet rather than an assistant. The products struggling hardest are the ones promising the most.

Where these companies actually are

Coverage of this category routinely describes any Chinese hardware startup as Shenzhen-based. It is often wrong, and it matters if you are sourcing, partnering or visiting.

Three of those are worth calling out because the error is so common. Ropet is repeatedly described as a Shenzhen company and is registered in Beijing. FoloToy and Fuzozo are both Shanghai. And TangibleFuture, maker of LOOI, is Dongguan rather than Shenzhen, which matters if you are trying to visit: Dongguan is roughly an hour out.

Two ways to build a face

Any robot that shows a human face uses one of two approaches, and they are not variations on a theme.

Rear-projected animated faces. A projector inside the head throws video onto the back of a translucent moulded mask. There are no facial actuators at all. Identity changes in software, or by physically swapping the mask in seconds. Furhat Robotics is the reference design, with a swappable polymer magnetic mask, speech in 40+ languages, camera tracking for up to ten people and a far-field mic array to about 5m. Engineered Arts builds the same idea into SociBot. The research lineage runs through Mask-Bot, Lighthead and Quori, and the papers are explicit that mechanical complexity stays low precisely because there are no actuators and no driven skin.

The weakness is light. Every rear-projected face has an ambient brightness above which it washes out, and it looks least convincing at oblique angles.

Actuated silicone skin. Micromotors pull anchor points in a cast silicone face. AheadForm's Origin heads run roughly 25 micromotors and 200+ control points with cameras behind the pupils, priced from about $50,000 to $150,000 by configuration. Hanson Robotics uses its proprietary Frubber elastomer, which folds more like skin than standard silicone. Realbotix builds face swapping in as a product feature with magnetically attached faces. Engineered Arts covers both camps: Mesmer for realistic skin, and Ameca, which deliberately declines to look human and sidesteps the uncanny valley by refusing to compete with it.

The weakness is service. Every one of those motors is a failure point, and a single dead actuator reads to a human observer as facial paralysis.

What a face costs to manufacture

If you are commissioning a silicone face rather than buying a finished head, you are buying three separate things and should insist they are quoted separately:

• The mould. Steel or aluminium, and the real capital cost. Ask how many pulls it is good for before fine detail degrades.
• The casting. Platinum-cure silicone, with shore hardness in the low teens for anything meant to read as skin. Hold a sample. The number on the datasheet does not settle it.
• The finishing. Multi-layer pigmenting and hair punching, which are manual, slow, priced by hour rather than part, and frequently subcontracted to a different shop than the one quoting you.

Geographically, the cluster is Dongguan around Songshan Lake, roughly an hour north of Shenzhen, where MoXian Technology among others works on silicone-based electronic skin with tactile feedback. Qingdao in Shandong is the other dense region. Shenzhen itself is where the brands, design houses and electronics are, not the silicone.

A warning on marketplaces: Alibaba and Made-in-China list many "humanoid robot silicone face" suppliers. Most are trading companies or doll factories rebadging catalogue moulds. The tell is whether they will quote a new mould with a tooling lead time or only offer existing faces.

How to evaluate one as a buyer

1. What happens if the company disappears? Local inference or an open SDK is the only answer that survives. Ask directly.
2. Is there a subscription, and what dies without it? If core personality is behind the paywall, price the subscription into the purchase over three years.
3. What is the single promise? Products that promise one thing outlive products that promise four.
4. Where does the audio go? For anything near a child, this is the question that matters most, and the 2025 incidents involving AI toys made it a regulatory issue rather than a preference.
5. Can you repair it? Ask about battery replacement specifically. A sealed battery sets a hard lifespan on an object designed to be loved.

Frequently asked questions

Are robot pets worth it for elderly companionship? The clinical evidence is strongest for simple, tactile, non-verbal devices rather than conversational ones. Products in the Moflin and Qoobo mould, which offer texture and responsive movement without language, avoid the confusion and the failure modes that talking devices introduce.

What happened to Moxie? Embodied announced closure on 26 November 2024 and the servers went dark on 30 January 2025, making an $800 device inert. Most buyers got no refund. Children lost a companion they had formed a real attachment to, which is why this is the category's most cited cautionary tale. A volunteer effort called OpenMoxie later revived units on local servers.

Is a $300 robot pet better than a $1,000 one? Frequently, yes. The cheaper products in this category tend to make one narrow promise and keep it. The expensive ones tend to promise general-purpose home robotics and miss.

Can I buy a robot with a realistic human face? Yes, but not at consumer prices. AheadForm's heads start around $50,000. For a face that swaps, the rear-projection route (Furhat and similar) is dramatically cheaper because it has no facial actuators.

Where are most of these made? Assembly and electronics in Shenzhen and the surrounding Pearl River Delta. Silicone skin specifically in Dongguan, concentrated around Songshan Lake.

Sources

Primary sources first, meaning the company's own documentation, the incident record or the paper. Secondary reporting follows, and is used only where no primary document exists.

Primary

1. Furhat Robotics, the Furhat robot - mask, projection, language and sensing specification.
2. Haivivi, BubblePal - the clip-on pendant product line.
3. KEYi Robot, Loona - Loona specification and pricing.
4. Living.ai, EMO - desktop pet specification.
5. Elephant Robotics, company - founding, location and the MarsCat, metaCat and metaDog line.
6. Ropet, about - Cute World Technologies, founders and founding year.
7. OECD AI Incidents, Embodied Moxie shutdown - the shutdown recorded as an AI incident.
8. Quori: a community-informed design of a socially interactive humanoid robot - the rear-projected face family, and why the mechanical complexity is low.

Further reading

9. The Robot Report, Anki addresses shutdown - the company's own statement on ongoing support.
10. VentureBeat, Anki shuts down after burning through almost $200 million - the raise and the closure.
11. AppleInsider, Moxie may be saved by a last-minute open-sourcing effort - the OpenMoxie revival.
12. TechNode, when AI companions break free from Wi-Fi with Fuzozo - Robopoet, Shanghai, and the cellular design.
13. Hackster, TangibleFuture's LOOI turns smartphones into companion robots - the phone-dock form factor.
14. Caixin, in depth: China's AI toy story - Haivivi in Qianhai, volumes and funding.
15. GDToday, Dongguan's Songshan Lake develops skin for robots - the silicone and electronic-skin cluster.
16. Live Science, Chinese company develops ultra-lifelike robot face - AheadForm Origin, actuator count and price band.
17. South China Morning Post, UBTech's lifelike humanoid robots built for companionship - the U1 consumer humanoid.
18. Furhat, overview - back-projection summary and history.