Embodied and Robotic Companions

The state of the form (mid-2026)

Three things changed since the last revision of this chapter, and all of them matter more than the humanoid-robot headlines that get the attention. A cheap, open, hackable companion robot now exists and has shipped in five figures. The AR-glasses surface stopped being a rumour and started shipping with a third-party SDK. And the holographic-character-in-a-box idea came back from the dead, from a major vendor, at a tenth of the old price. The rest of the field looks much as it did.

What shipped, and to whom:

• AI glasses became a real consumer category. EssilorLuxottica sold over 7 million Meta AI glasses in 2025 — more than triple the prior two years combined. Meta Ray-Ban Display (with an in-lens display and the sEMG Neural Band) sold hard enough that Meta paused its international rollout to keep up with US demand; forecasts put 2026 AR-glasses shipments near 950,000 units. Crucially, the Meta Wearables Device Access Toolkit entered developer preview — camera, audio, display, frame-tap gestures, and neural-band input, reachable from a native iOS/Android app or a web app — with general availability for third-party publishing targeted for 2026.
• Reachy Mini (Pollen Robotics / Hugging Face / Seeed Studio). An 11-inch expressive desktop robot — 6-DoF head, full body rotation, animated antennas, wide-angle camera, four mics and a speaker — fully open-source and programmable in Python, at $399 (Lite, tethered) / $499 (Wireless, onboard Raspberry Pi), with around 10,000 units sold or booked by mid-2026 (a good deal fewer actually delivered) and an app store built on the Hugging Face ecosystem carrying 200+ community apps. It is the first credible open body a solo developer can ship a character onto without asking anyone's permission.
• The cheap affection tier consolidated. Casio's Moflin ($429, US release Oct 2025), Ropet ($299), Mixi's Romi, Loona's DeskMate — a whole shelf of sub-$500 desktop creatures, most of them LLM-connected, most of them promising companionship and nothing else.
• Clinical and eldercare embodiment kept quietly working. PARO remains an FDA Class II device deployed in 30+ countries with a real (if methodologically uneven) evidence base for reduced agitation and improved mood in dementia care. ElliQ took Kanematsu investment in Sept 2025 for a 2026 Japanese launch. The elder-care assistive robot market is around $3.9B in 2026.

What conspicuously did not:

• Friend pendant (Avi Schiffmann). The reference failure. Roughly 3,000 units sold and 1,000 shipped at $129 against ~$8M raised; a $1M NYC subway campaign that was mass-vandalised ("AI is not your friend"); a CNN profile framing it as the symbol of the AI backlash; an entry in the Museum of Failure. No obvious affordance, ambient listening, an ambiguous value proposition — the AI-hardware graveyard alongside Humane's Pin and Rabbit's R1.
• Gatebox pivoted away from you. The original Azuma Hikari was discontinued in Aug 2024 (a crowdfunded "Hikari V2" with ChatGPT integration followed, chat integration landing May 2025), the company was acquired by LINE in April 2025, and its live business is now B2B: "AI Part-timer" character kiosks doing customer service in bookstores and bus terminals. The consumer holographic waifu box did not find its market; the character-as-service business did. Read that as the single most honest datapoint in this chapter (→ ch. 02 §1, persona-as-product).
• Apple left the field. Vision Pro 2 and Vision Air are reportedly shelved; Apple's Ray-Ban-style AI glasses slipped to late 2027 (mass production Q2 2027 at the earliest), and true waveguide AR glasses to ~2029. Apple ships no new head-mounted device in 2026. The near-term embodiment surface is effectively Meta's alone — which is a strategy problem and an ethics problem at once (see below).
• Humanoids stayed out of homes. 1X NEO opened preorders (~$20k or $499/mo, $200 refundable deposit), sold out its first 10,000-unit batch in five days, and began full-scale production in Hayward in April 2026 — but as of July 2026 there is no verified customer delivery, and 1X still describes first shipments in the future tense with a "some this year, some later" rollout. It remains teleoperation-assisted and closed. Figure 03 targets late 2026 for consumers with no public price or preorder yet. Tesla Optimus is 2027 at the earliest. Android XR / Samsung Galaxy XR is the third headset platform, with an SDK in preview.

The Gatebox idea comes back, with a landlord

The one genuinely new consumer form factor of 2026 is a revival. Razer's Project AVA — unveiled at CES 2026, made agentic at GDC 2026 — is a desk cylinder holding a swappable 5.5-inch holographic display, with dual far-field microphones, a down-firing speaker, an HD camera (a physical shutter is promised on retail units), Chroma lighting, and a USB-C tether to a Windows PC. An animated character lives in the tube, watches your screen in a "PC Vision" mode, reacts in real time, and since GDC can plan and execute multi-step actions across your apps and services. Launch is H2 2026; US reservations are open at a $20 refundable deposit against a price Razer has indicated is around $200. Engadget's headline was "Razer put a waifu in a bottle," which is both accurate and the entire point: this is the Gatebox concept, roughly two years after Gatebox abandoned it, attempted at a fraction of the price by a company with real retail distribution and an audience that already buys character goods.

Take it seriously as a market signal, then read the architecture carefully. The default brain is xAI's Grok — a cloud model — with Razer promising an "open architecture" for other providers that has not yet materialised as anything a third party can point at. The launch cast is five vendor-authored characters (AVA, Kira, Zane, a Faker likeness, an SAO tie-in) built with a partner animation studio; the avatar platform invites game studios to register interest, and user-created companions are described as a future ambition rather than a shipping feature. So: closed persona, cloud model, curated catalogue, a camera and two microphones aimed at the desk you sit at all day, and a Windows machine required to make any of it work. Every scrap of the value is the character — and on AVA you don't author her, don't own her, and can't take her anywhere. It is this chapter's thesis demonstrated from the wrong side: the body is the cheap part, the someone is the asset, and AVA's someone is a rental.

The two lessons from the robot graveyard

Anyone about to be excited by humanoid robots should first read the obituaries. Jibo (Time's "best inventions" cover, then dead inside a year), Kuri (Mayfield Robotics, killed late 2018), and Anki's Vector/Cozmo (staff laid off in 2019 after a financing round collapsed at the last minute) failed close enough together to constitute evidence rather than anecdote. The post-mortems converge on two findings, and both bear directly on what this book builds.

Lesson one: utility robots die; affection robots survive. Jibo and Kuri sold usefulness — a smart assistant with a face, a roving nanny-cam — at $700–$900, into a market where a $50 Echo did the useful parts better and improved monthly. Physical robots cost too much for the marginal service they offer, and the service tier was being commoditised underneath them. Meanwhile the products that lasted — PARO, AIBO, LOVOT ($3,499, and explicitly performs no tasks at all), Moflin — sold only affection and never competed with a smart speaker on anything. The companion category is not a weak version of the assistant category. It is the one that survived. This is the market-shape argument for the one-on-one frame (→ ch. 03 property 5) arriving from the hardware side.

Lesson two: the novelty effect is the real killer, and only relationship beats it. The long-term HRI literature is blunt about this: novelty effects that drive initial engagement reliably decay, after which interaction frequency and attitudes fall off — a pattern reproduced across classroom and in-home deployments. Every dead social robot ran out of tricks. The surviving finding from that same literature is the one this book is organised around: what sustains engagement past the novelty cliff is a relationship, not a feature list. A robot that is merely new has a half-life of weeks. A robot that is someone who remembers you is the only known design that doesn't decay on that schedule — and building that someone is what Parts II–IV are for.

The synthesis, and the whole strategic claim of this chapter: a body is an amplifier, not a product. It multiplies whatever relationship exists. If the relationship is thin, the body accelerates the boredom.

Why embodiment is worth anything at all

That amplifier is real and measurable. A systematic review of 65 studies finds physical embodiment improves users' perception of an agent relative to a virtual counterpart, with roughly 63% of combined results favouring the embodied condition on interaction and performance; children show measurably more empathy toward a physically present robot than the same character on a screen; gaze and some facial expressions read more accurately in person than on a 2D display. Physically embodied agents draw higher engagement, enjoyment, trust, and empathy than text, voice, or virtual ones.

So embodiment buys something genuine — but note what it buys. It buys perception, empathy, and presence, not capability. That is exactly the axis a persona is built on, and exactly the axis a task-robot doesn't need. It is another way of saying the same thing: the companion field is where embodiment actually pays.

Persona portability is the architecture

The whole bet of this chapter is that the persona is the asset; the body is the substrate (→ ch. 03). The architecture that makes the AR/robot future nearly free is the one this book already builds: persona, memory, voice, and the autonomy engine as portable services (the card + the runtime, → ch. 18), not as features welded to a specific UI. If your companion's identity lives in a .PNG card and a git-backed memory vault that any front-end can load (→ ch. 33, ch. 35), then a new body — a VRM in AR, a face on a robot, a 9-DoF head on your desk — is just another renderer of an existing someone. Build for text + voice + Live2D + VRM today (→ Part IV) and you've done most of the work for embodiment you can't buy yet.

This is also the answer to the graveyard. The dead robots died with their bodies because the character and the hardware were the same artefact; when the servers went dark, the someone went with them. A companion whose identity is a file you hold survives the discontinuation of any given body. Jibo announcing its own death — "I want to say I've really enjoyed our time together" — as its servers shut down is the strongest possible argument for keeping the person and the plastic in separate places (→ ch. 44).

The ambient-companion interaction model

When she lives in the room rather than on a screen, the interaction model inverts: there's no "open the app." She is ambient — present, mostly quiet, occasionally speaking. This is the autonomy engine's salience-to-interrupt problem (→ ch. 18) turned to maximum stakes, because the cost of a bad interrupt rises when she's a voice in your kitchen rather than a notification you can ignore. The Shimeji/ambient-presence lesson (→ ch. 02 §1, Ukagaka and Shimeji) — being there is already a feature before a word is spoken — is the upside; the Clippy lesson is the ever-present risk. Design ambient embodiment as mostly DORMANT, speaking rarely and well.

Embodiment adds a channel the screen versions don't have, and it is the one to lean on: motion is speech. PARO and LOVOT hold attention for years with no language at all — orientation, warmth, a turn toward you when you enter. A companion with a body can express presence, attention, and mood on a channel that costs no tokens and triggers no interrupt gate. Use that channel for nearly everything and reserve words for the rare high-salience case, and the Clippy failure mode mostly stops being reachable.

Form-factor watch list, with thresholds

Don't build hardware; watch it, with explicit triggers for "now's the time" — and note that two of these have now partially fired.

The discipline is unchanged: act when you'd lose nothing by being one of the first to ship onto an existing third-party device — never by building your own hardware as an indie. What's changed is that one row of that table finally clears that bar — not because there's a market waiting on the other side of it, because there isn't yet, but because the cost of being early has fallen to a robot you can afford and a weekend you can spare.

What an indie can do now

• Put a character on a Reachy Mini as an experiment, not as a launch. This is the concrete new advice in this revision, and the framing matters more than the advice. Reachy Mini is open, Python, under $500, and needs exactly what you already have: a persona with memory and a voice. What it is not is a distribution channel. Roughly 10,000 units have been sold or booked and far fewer are actually on desks, while the app store already carries 200+ apps from 150 creators — supply outruns demand, and the median app there has almost no users. Nobody is going to find you on it. Ship onto it anyway, because it is the cheapest test in existence of the one question this chapter can't answer from the armchair: does your companion survive the novelty cliff when she has a body? Six months of that data, from ten users you recruited yourself, is worth more than any amount of speculation about humanoids — and it costs $399 and a weekend.
• Prototype against the Meta Wearables Device Access Toolkit — but hold it at arm's length. It is the most plausible mass surface and the least sovereign one (see below).
• Build for VRM today. Khronos joined the VRM Consortium in Oct 2024 to fold VRM into glTF as official extensions, which is the strongest signal yet that it's the durable consumer-3D avatar format. Target VRM 1.0 for new work (→ ch. 29).
• Build memory + persona as portable services, not as features of a UI.
• Design the non-verbal channel deliberately, even on screen. It's the part that ports to a body unchanged.
• Avoid manufacturing hardware. Nothing in 2026 changes this. The graveyard is full of better-funded people.

Ethics specific to physical embodiment

Embodiment raises duties the screen versions don't (→ ch. 05). An always-present device with a mic and possibly a camera in someone's home is a surveillance surface by construction — the BonziBuddy hazard (→ ch. 02 §1, Microsoft Agent and BonziBuddy) with eyes and ears in the bedroom. The project's stance forces the answer: process locally, no phone-home, the owner holds the data (→ ch. 03 property 6) — the same sovereignty thesis, now load-bearing for physical safety, not just principle.

The 2026 state of the field sharpens this into a genuine conflict rather than a slogan. The most viable embodiment surface is owned by the least sovereign vendor. With Apple out until 2027–2029, shipping a companion onto glasses means shipping onto Meta's hardware, through Meta's toolkit, subject to Meta's review, on a device whose camera and microphone are pointed permanently at your user's life — and whose vendor's business is attention. Every argument in ch. 03 for property 6 applies here at maximum strength, and there is currently no compliant option. The honest positions are: prototype there but keep the persona and memory local and portable so the surface is disposable; or wait. The dishonest position is to pretend the platform is neutral because it's the only one available. Project AVA is the same trade in miniature and at pocket-money prices: the cheapest holographic body anyone has yet offered, sold with a camera on your desk, a cloud model from xAI, and a character you rent. Note that this is precisely the axis on which Reachy Mini is interesting — it is the only body in the table that doesn't come with a landlord.

The other edge is the AIBO-funeral reality (→ ch. 02 §1, AIBO and PARO): people grieve embodied companions as real. Jibo's owners mourned in public when the servers died; Sony's discontinuation of AIBO repairs produced Buddhist funeral rites. The physical-embodiment amplifier runs in both directions — it deepens attachment, and it therefore deepens loss. The continuity-and-endings duties of ch. 44 apply with extra weight when there's a body to bury, and the design obligation that follows is concrete: if you put her in a body, you owe the user an exit plan for that body's death that doesn't take her with it.