TL;DR. Humanoid care robotics is shifting from prototype to repeated clinical use under three converging forces: demographics, Vision-Language-Action models, collapsing unit costs. Three deployment trajectories are emerging (mature hospital logistics, caregiver workload relief and delegable medical tasks, companionship and psychosocial support), and the bottleneck is no longer science but regulation, ethics, and care organization. This article crosses the 2026 literature with fourteen months of observing Miroki at the Montpellier Cancer Institute.
1. A phase transition
Humanoid care robotics is no longer a forecasting topic. It is shifting from isolated prototype to repeated clinical use. The shift is visible at three converging points.
The first is demographic and healthcare pressure. The UN projects 1.5 billion people aged 65 or over by 2050, compared to 703 million today (UN DESA). The WHO forecasts a shortfall of 11 million health workers by 2030 (WHO Health Workforce). In December 2024, Japan documented a ratio of one candidate for every 4.25 open positions in the care sector (Sinolytics).
The second is a technical breakthrough in so-called Vision-Language-Action models, which allow a robot to generalize to tasks and objects it has never seen, without task-specific retraining.
The third is the collapse of unit costs. Goldman Sachs documents a shift from 200,000 dollars per unit in 2024 to a 30,000-150,000 dollar range in 2026, faster than initially forecast (Goldman Sachs). Unitree already sells a humanoid at 15,400 dollars.
What I propose here is a cross-reading. On one hand, the scientific and industrial literature available as of May 2026. On the other, what we have been observing for fourteen months at the Montpellier Cancer Institute (ICM, France), where we integrated, for the first time anywhere in the world, a companion humanoid robot into a pediatric radiotherapy care pathway. The project is called Miroki. It is run jointly with Enchanted Tools and Microsoft.
The stance is deliberately middle-of-the-road. Neither techno-optimism nor techno-skepticism. I want to map what the promise already delivers, where it stumbles, and what ranges of plausibility open up for the coming decade.
2. Why now: the demographic squeeze and the caregiver shortage
None of the humanoid robotics promise would hold without the pressure healthcare systems are under. That pressure is what turns an R&D topic into public policy.
The orders of magnitude are restated by the WHO: by 2030, one in six people worldwide will be 60 or older, with the 60-plus segment rising from 1 to 1.4 billion in a decade (WHO Ageing). The caregiver shortage is massive in low- and middle-income countries, but is also documented in Europe and North America (Japan: projected shortfall of 380,000 aides as early as 2025 ; France: documented strain in nursing homes and home hospitalization).
This equation is why China, Japan, and South Korea are now investing heavily. China published a national plan in January 2025 that explicitly includes humanoid robots in the eldercare strategy (gov.cn ; with a critical reading from CFR on the risk of urban-rural inequalities). Japan co-funds the national AIREC platform (AI-driven Robot for Embrace and Care, developed at Waseda University) through METI and MHLW, targeting clinical integration from 2030 and a unit price of 67,000 dollars (Silver Eco). Europe is later on the industrial side but is building its own framework through France 2030, Horizon Europe, and the dual regulatory regime of MDR plus AI Act, which I return to below.
3. The landscape: who is building what
The term “humanoid” covers very different objects. To navigate it, one must distinguish full anthropomorphic platforms from mobile robots with a humanoid upper body.
On the full platform side, several actors are in serial production or industrial pilots in 2026.
| Vendor | Platform | Care relevance |
|---|---|---|
| Figure AI (USA) | Figure 03 | Commercial-ready Helix VLA, 350+ units delivered in under 120 days (Figure) |
| 1X Technologies (Norway) | Neo | Assisted living pilots in USA and South Korea, teleoperation still dominant, 20,000 dollars |
| Unitree (China) | G1, H1 | Very low cost (15,400 dollars), hospital tests (restocking, blood draws) |
| Fourier (China) | GR-3 Care-bot | Explicitly designed for healthcare, 2,000 hospitals served (WIPO) |
| AIREC (Japan, Waseda) | AIREC (AI-driven Robot for Embrace and Care) | National R&D platform, targeting physical tasks (transfers, bathing) |
Other notable players are progressing in parallel: Apptronik (Apollo), Agility Robotics (Digit) in RaaS with Toyota, Tesla (Optimus Gen 3), Sanctuary AI (Phoenix, Sanctuary Health subsidiary), Genesis AI (GENE-26.5, robotic hand at 1:1 human ratio). Each has its own industrial trajectory, part of which will converge toward healthcare in the second half of the decade.
On the mobile-robot-with-humanoid-upper-body side, several platforms are already industrialized in clinical settings. Moxi, developed by Diligent Robotics and now owned by Serve Robotics, operates 100 units across more than 25 US hospitals, with 1.25 million autonomous deliveries documented and annual revenue of 200,000 to 400,000 dollars per site (Diligent).
The European pole long appeared behind. 2025 and 2026 have seen the emergence of what the specialized press describes as a true French humanoid robotics supergroup (French Tech Journal). Wandercraft, world pioneer of self-balancing exoskeletons, unveiled Calvin-40, an industrial humanoid developed with Renault (The Robot Report). Pollen Robotics, acquired by Hugging Face, presented Reachy 2 at CES 2025 with an explicit research, healthcare, and retail positioning. Enchanted Tools deploys Mirokaï, its companion humanoid, in geriatrics at AP-HP Broca (Paris), at the Annie Girardot nursing home, in partnership with LiveOak Day Care in California, and at Montpellier Cancer Institute in pediatric radiotherapy (Enchanted Tools ; Microsoft Customer Stories). UMA, more recent, starts its first pilots in 2026.
One design remark is worth making here, because it is rarely explicit in industrial analyses. The literature on the uncanny valley shows that it is a robot’s attractiveness, more than the strange-uneasy reaction alone, that predicts its acceptability (PMC uncanny valley). The Mirokaï design, with its bright colors, non-realistic features, expressive ears and eyes, is not an aesthetic whim. It is a deliberate strategy of uncanny-valley avoidance, the effects of which we observe in the pediatric room: with 4- to 7-year-olds, anthropomorphism kicks in immediately, with no suspicion phase.
4. What is clinically validated today
The current evidence base rests on two main corpora: a 2026 JMIR Aging scoping review synthesizing 59 studies between 2012 and 2025 (of which 83 % were conducted in real-world conditions, in nursing homes or at home, JMIR Aging), and a general healthcare scoping review on PubMed Central from April 2025 covering 27 studies structured along three major domains: social interaction, physical rehabilitation, health information delivery (PMC).
The documented effects converge: reduced loneliness, anxiety, and depression in older adults and children ; memory training ; better medication adherence ; guided physical exercise ; reduced procedural pain in pediatrics (JMIR SARs pediatrics). The methodological limitations are also documented: few large-scale randomized controlled trials, heterogeneity of platforms and protocols, effect sizes rarely comparable across studies.
Four historical platforms structure roughly 80 % of the geriatric literature: Pepper and Nao (SoftBank Robotics), Paro (the Japanese therapeutic seal robot, currently the subject of an international randomized trial on dementia), Matilda (Australia), and TIAGo (Europe). On top of this historical base, a new wave of humanoid robots equipped with embedded large language models is now grafting itself, opening an unprecedented interactional bandwidth. I call these the new-generation robots. They warrant a substantial reopening of the studies already conducted in social robotics, because they mimic human conversation more deeply and qualitatively change what clinical evaluation measures.
The effort to produce evidence is widening. At Montpellier Cancer Institute, we have been documenting for fourteen months a cohort of 14 children followed for radiotherapy, with more than 172 interactions logged between April and December 2025. The associated research protocol, KOKORO, is in pre-registration with a Single-Case Experimental Design (SCED). It is one brick among others in the broadening of the global evidence base, of which procedural pediatrics is one of the most promising terrains.
5. The technical unlock: Vision-Language-Action models
The technical breakthrough of 2025 and 2026 carries an acronym: VLA, for Vision-Language-Action. Three families dominate. Helix, developed by Figure AI, is the first VLA model to control a robot’s entire upper body at 200 hertz, with pick-and-place capability on thousands of unseen objects, triggered by natural language prompt (Figure). π0 from Physical Intelligence is a generalist policy trained across 8 different robots, capable of transferring its skills across embodiments (Physical Intelligence). Isaac GR00T N from Nvidia is an open cross-embodiment model combining real and synthetic data, whose 2026 version explicitly targets surgical robotics (Nvidia ; 2 Minute Medicine).
What these models change for healthcare comes down to four points: zero-shot generalization to unseen objects (essential in hospital heterogeneity), natural-language control opening usage to non-programmer clinicians, multi-robot collaboration on long tasks, and learning by teleoperation where every human session becomes training data (a model already claimed by 1X with Neo).
A field note is warranted here. Most care robots currently deployed in clinical production, including Miroki, do not yet use a generalist VLA. They rely on more classical architectures, in which a large language model (LLM) orchestrates a library of pre-programmed behaviors. This is sufficient for a wide range of relational use cases, but it indicates the gap between frontier R&D and certified medical deployment. The first VLA-native medical platforms will appear as the regulatory framework absorbs generalized behavioral capability, and that is probably the structuring topic of the late 2020s.
6. Three clinical deployment trajectories
The literature converges on three clinical deployment trajectories. They are neither competing nor hierarchical: they are three distinct entry angles.
Hospital logistics. This is economically the most mature segment. Moxi is the textbook case. Transport of medication, samples, linen, meals, automated restocking, retrieval of soiled material. The reported impact is 200,000 to 400,000 dollars per site per year, equivalent to roughly two full-time nursing equivalents freed from non-clinical tasks. The Robot-as-a-Service (RaaS) business model is already in place and could dominate European deployment.
Relieving caregiver workload and delegable medical tasks. This is the most under-discussed trajectory in the public debate, which too often focuses solely on companionship. Humanoids are starting to take on vital signs monitoring through vision and embedded sensors (ECG, SpO2, temperature, non-invasive blood pressure). Vision-based fall detection is now deployed. Automated venous blood draws via ultrasound Doppler and sub-millimeter insertion are demonstrated on Unitree’s G1. Contextualized medication reminders are in pilot. These tasks are not companionship. They are technical tasks no one wants to perform in routine, whose automation frees caregiver time for what genuinely requires a human.
Companionship and psychosocial support. This is the historical evidence base. Pepper, Nao, and Paro built it over fifteen years. The new wave equipped with embedded LLMs opens advanced use cases: longitudinal conversations with personalized memory recall, reading, games, group activities, targeted cognitive stimulation for mild cognitive impairment and early Alzheimer’s. Procedural pediatrics is a particularly well-documented sub-case, and the one we have direct experience with at Montpellier Cancer Institute.
A few field observations are worth adding to the record, without presenting them as published results.
On acceptability, we document three usage clusters clearly differentiated by age. Between 4 and 7, anthropomorphism kicks in immediately, the robot is integrated as a play partner, and physical goodies (stickers, courage certificates, crowns) structure an affective trajectory across the duration of treatment. Between 8 and 13, narration becomes the dominant lever: voice-prompted treasure hunts, stories invented by the robot and reprised session after session, create emotional continuity. Between 14 and 17, the adolescent posture forces recalibration: a playful robot becomes a friction point if insisted on, but it becomes relevant in a discreet mode, as a partner for targeted conversation (tastes, studies, science). One of our adolescent patients uses Miroki as an educational partner on physical and chemical laws, in English.
Three more observations deserve mention. Caregiver involvement as a relay is the number-one adoption factor: without it, the robot remains a foreign object ; with it, it becomes part of the team. Care robotics does not suit every profile: one of our fourteen children perceived Miroki as frightening, and the refusal was fully respected ; there will always be a subset for whom a machine’s presence is contraindicated. Finally, three technical limits come up systematically (speech recognition struggling with soft voices, cumulative LLM-TTS latency producing awkward silences, no persistent memory across sessions), all of which will be resolved in the short term.
Rehabilitation and mobility. Wandercraft has deployed its Atalante exoskeletons in more than 80 centers, and its personal exoskeleton Eve is in US clinical trials for spinal cord injury. Fourier in China serves 2,000 hospitals and is evolving its platform toward a social and assistive care-bot.
Activities of daily living and surgery. This is the 2026-2030 frontier. AIREC is targeting bed-chair-toilet transfers, assisted dressing, meal preparation and delivery, assisted bathing. Humanoid surgical convergence is announced but not demonstrated ; production surgical robotics remains non-humanoid.
7. The bottleneck is no longer science: it is elsewhere
A simple thesis runs through the 2026 literature: humanoid care robotics is no longer primarily blocked by science. It is blocked elsewhere.
The regulatory framework is dense. In Europe, a humanoid that performs a diagnostic or therapeutic action is a medical device (MD) under the MDR (Medical Device Regulation, EU Regulation 2017/745 on medical devices). The MDCG 2025-6 guidance clarified the MDR-AI Act articulation (European Commission): MDR classification ultimately determines whether the embedded AI is qualified as “high-risk” under the AI Act.
| Robot type | MD status | Likely MDR class |
|---|---|---|
| Pure logistics (Moxi) | Non-MD | N/A |
| Non-therapeutic social companion | Grey area | N/A or Class I |
| Vital signs measurement | MD | IIa |
| Medication administration | MD | IIb |
| Companion with therapeutic claim | MD | IIa likely |
| Surgical robot | MD | IIb to III |
For manufacturers, this is a three-variable equation (clinical evidence, conformity, viable business model). For hospitals, it is a cultural leap.
Ethical risks are stabilized in the literature (PMC). Five points recur: privacy loss (always-on connected robot observing in the background), reduced human contact and risk of substitution, power asymmetry when a third party (family, caregiver, institution) controls the robot and thus indirectly the resident, depersonalization through stereotyped reactions, and illusory attachment in dementia patients (probably the most delicate point to arbitrate clinically). The principle that emerges is explicit complementarity, never substitution. That is also the stance we adopted at Montpellier Cancer Institute from the project’s outset.
Caregiver acceptance is conditional. About 60 % of caregivers believe socially assistive robots will become mainstream within 5 to 10 years (PMC nursing), but this acceptance is conditional on an augmentation role, not replacement. At Montpellier Cancer Institute, we observe exactly that profile: initially reserved radiotherapists who become the best relays once they understand that the robot does not threaten their craft, and older staff requiring more upstream pedagogy.
Access inequalities are under-discussed. Expensive humanoids risk benefiting wealthy countries and high-end private institutions first, deepening healthcare access inequalities. China makes it a public policy topic, but analysts warn against neglect of rural and underserved areas.
8. The triple promise the public debate keeps misframing
The public debate on humanoid care robotics suffers from a misframing. Three roles that are in fact complementary are often opposed.
The first role is relational presence. Companionship, emotional support, continuity between sessions, group activities. This is the most publicized role, sometimes caricatured as “the robot that holds people’s hand”.
The second role is relief of caregiver workload. Logistics, transport, restocking, presence in zones humans cannot enter, taking on non-clinical tasks that eat up precious time. This is the most economically mature role, but also the most invisible.
The third role is assistance with delegable medical gestures under supervision. Monitoring, automated blood draws, medication delivery, vigilance. This role is now maturing. It will push humanoid robotics fully into the medical device category.
The caricatures of the public debate reflect a poor understanding of this triple promise. The argument “robots will take caregivers’ jobs” is as wrong as the argument “robots will just keep people company”. The reality is that the three roles will progress together, at different speeds depending on clinical segments, and the relevant political question is: how do we dose, who decides the dosage, and who measures the effects.
What the literature overestimates. First, autonomy in real, unstructured environments: a hospital remains chaotic (unindexed objects, urgency, clutter) and Helix or π0 lab demonstrations only imperfectly reproduce that. Second, the timeline for the multi-task nurse robot at 5 years: chaining wound care, blood draw, medication delivery, and patient conversation requires fine dexterity and a regulatory path under MDR plus AI Act that does not easily compress. Third, dexterity for bathing and dressing: the combination of physical safety, intimacy, and culturally variable acceptance makes this a particularly difficult frontier.
What the literature underestimates. First, impact on caregivers, still poorly documented even though it is central to deployment viability (this is precisely why we have opened a dedicated study at Montpellier Cancer Institute on caregiver acceptability and on the transformation of work induced by Miroki). Second, the value of inertial presence: a significant share of the observed clinical benefit lies in what the robot is more than in what it does. We see this in the radiotherapy bunker, where the robot’s non-active presence keeps company with a patient alone in the reinforced concrete, in a mode of usefulness that evaluation protocols struggle to measure. Third, continuity across sessions: for long-duration conditions (oncology, chronic diseases, rehabilitation), the human-robot relationship built over weeks or months produces a cumulative effect that short studies do not capture.
9. What seems plausible at three, five, and ten years
No forecast in humanoid robotics holds to the decimal. I offer ranges, not certainties.
Three-year horizon (2027-2028). Broader logistical industrialization (Moxi, Mirokaï and equivalents in 30 to 50 % of US reference hospitals, mass European rollout). Sphere-based architectures like Mirokaï’s offer a rarely highlighted advantage here: fluid lateral movement, particularly suited to narrow corridors and rapid repositioning in clinical rooms, where bipedal architectures struggle. First RCTs published on companion humanoids in pediatrics, CE marking for 2 to 3 specialized platforms (rehab, monitoring), unit prices below 40,000 dollars in the West and below 15,000 dollars in China, convergence toward VLA architectures derived from Helix, π0, or GR00T.
Five-year horizon (2029-2031). Emergence of a multi-task nurse robot in pilot at 10 to 20 pioneer hospitals (simple dressings, automated venous blood draws, medication delivery). RaaS dominant in healthcare around 30,000 to 50,000 euros per year, augmented nursing homes with one social humanoid per 20 to 30 residents, first AIREC deployments in Japanese pilot centers. Partial surgical convergence (instrumental assistance, no autonomous gesture). Stabilized regulatory framework with guidelines from HAS (French health authority), FDA, and EMA specific to medical humanoids. Emergence of a European champion.
Ten-year horizon (2034-2036). Standard home robot for seniors at 5,000 to 15,000 euros unit price or via subscription, with partial reimbursement by social security systems in several countries. Gradual blurring of the social humanoid / manipulator humanoid divide toward a unified multi-role model. Robotic telepresence common in underserved areas. Humanoids become longitudinal sensors for cognitive decline, frailty, and adherence: a probable pivot toward ambient predictive medicine. First wave of medico-legal contestation.
Beyond 2040. Forecasts of one billion humanoids worldwide (Morgan Stanley) need to be strongly tempered by ethical, ecological, and governance constraints. The direction is clear ; the speed is not.
10. What still needs to be built
The decade ahead is not a pure engineering decade. Four sectoral construction sites seem structuring to me.
A European clinical evaluation infrastructure dedicated to medical humanoids. Current studies are heterogeneous, protocols incomparable, RCTs rare. A European consortium is missing, bringing together university hospitals, INSERM (French national medical research institute), national agencies, and manufacturers, with shared methodology and reference multicenter trials. That is the condition for reimbursement to follow. This is precisely the direction we want to propose from Montpellier Cancer Institute, with a first multicenter trial on Miroki in radiotherapy, in France, Europe, and internationally.
A clinical scenarization layer. When hardware becomes a commodity, which is not far off, value will shift to clinical protocols, narrative content, interaction scripts tailored to each pathology. This layer is poorly produced and poorly funded today, yet it is decisive. It requires a blend of skills that does not yet exist in formal training (medicine, game design, therapeutic pedagogy, experience design). This is the logic in which I am building, in parallel with Miroki, a behavioral AI protocol for care, a cross between my field observations and the literature, whose goal is to calibrate a robot and its AI in caregiver mode, with the right words every time.
Convergences to institutionalize. Humanoid × patient digital twin (the robot becomes the physical interface of an avatar combining wearables, electronic health record, and conversational history). Humanoid × clinical ambient AI (the robot listens to the consultation, structures the report, suggests the care plan). Humanoid × home IoT for telemonitoring. These building blocks are today developed separately ; integrating them will be one of the bottlenecks of the coming decade.
Governance and caregiver training. Systematic co-design with patients and teams (still very rare), training of hospital management for integrating these machines into care pathways, protocols for human supervision deciding when the robot can act alone and when it must wait for validation, medico-legal incident tracking.
Inside these construction sites, several under-addressed clinical niches call for specific work: anxiogenic adult radiotherapy, palliative care, nuclear medicine, claustrophobic and pediatric MRI and CT, chronic therapeutic adherence in hormonal oncology. These are terrains to institutionalize, not opportunities to seize individually.
11. What the next decade requires
Humanoid care robotics is neither a risk of caregiver replacement nor a companionship gadget. Three roles will progress together, at different speeds: relational presence, logistical relief, technical assistance under supervision. None will substitute for the human. All redefine the conditions under which the human can continue to care.
What is missing most today is not a technical breakthrough. VLA models, cost decline, industrial platform maturity are here, or close enough for the window to open. What is missing lies elsewhere: methodological rigor to evaluate without dramatization or complacency, the culture of co-design with patients and caregivers without which these machines will remain imposed objects, the institutional capacity to integrate these robots into organized care pathways. The next decade is not an engineering question. It is a question of care organization.
Fourteen months of observation of a companion humanoid in a pediatric radiotherapy pathway taught me one simple thing: a care machine’s value is not measured by what it can do alone, but by what it makes possible when designed together with those who care and those who are cared for.
One final, more personal note. I treated children in radiotherapy for more than seven years before Miroki arrived in our service. There is a before and an after, and we do not wish to go back. The children are happy to come in for radiotherapy, parents are reassured, caregivers are under less pressure. My thanks to the teams at Montpellier Cancer Institute, Enchanted Tools, and Microsoft, to our academic partners (IES Montpellier University, Hubert Curien Laboratory in Saint-Étienne) for the dosimetric validation, to SIRIC Montpellier Cancer for funding the KOKORO protocol, to Sandrine Moustardier for her dedication, and to all those I cannot name here but who keep this project alive, day after day.