China is turning factories into schools for robots. Tesla is betting on general-purpose labour. But Physical AI’s most urgent value may be filling the gap left by a simple human limitation: no one can always be there.

Sinclair

This is not an abstract technology question for me.

Over the past six months, my mother, who is in her nineties, has fallen several times at home. My sister and brother-in-law care for her, and we take her back to the hospital for regular checkups. The family is doing a great deal. But no one can stand beside another person twenty-four hours a day.

Thankfully, none of the falls caused serious injury. They did, however, leave me with a question I could not ignore: Would the outcome have been different if a physical AI system had been continuously present — watching for risk, offering a timely reminder, and calling for help the moment something went wrong?

I am not imagining a machine that replaces the family. I am imagining a system that fills the gap created by the fact that human beings cannot always be present.

That is also where I began to see Physical AI differently. Its deepest value may not be making robots dance, run, or perform on stage. Nor is it merely the next capital-market narrative. The more important question is this: when AI leaves the screen and acquires a body, can it perceive risk, respond in the physical world, and bring a real person back into the loop when human presence is missing?

Most of AI’s recent achievements have taken place behind a screen. It answers questions, writes code, generates images, and analyses information. But for an older person getting out of bed at night, even the most capable chatbot remains in another world if it cannot see her lose balance, check whether she is hurt, or alert someone after she falls.

AI is now beginning to leave that world.

In factories, workers are asked to move materials, sort parts, assemble products, and work continuously. In the home, it is imagined as a general-purpose assistant. In China’s emerging market for lifelike humanoids, it is even being sold as a companion with expressions, memory, and a persistent personality.

When AI leaves the screen, will people first pay for its labour, for the safety of someone they love, or for the feeling that another presence is in the room?

1. What Does Physical AI Actually Add?Physical AI, embodied intelligence, and spatial intelligence are often treated as interchangeable. They overlap, but they are not identical.

Spatial intelligence, the field Fei-Fei Li has recently emphasised, concerns an AI system’s understanding of the three-dimensional world: depth, geometry, motion, object relationships, and how things interact in space. Embodied intelligence goes one step further. Intelligence is no longer confined to reasoning inside a model; it learns through a body acting in an environment.

Physical AI is the broader industrial expression of this idea. It connects perception, spatial and physical reasoning, planning, control, and action into a deployable loop. Its embodiments include humanoid robots, autonomous vehicles, drones, industrial machines, and intelligent spaces.

The generative-AI loop is roughly:

Prompt → model reasoning → text, image, or code

The Physical-AI loop is different:

Sense the world → estimate physical state → plan → act → observe contact and outcome → detect error → correct

Large language models learn primarily from records humans have already produced: text, images, code, and other digital traces. Physical AI must learn from another kind of data. It has to know not only what the word “cup” means, but where the cup is, what it weighs, where it can be grasped, and what may happen if the fingers apply too much or too little force.

Its unit of experience is not merely a sentence or an image. It is a perception–action–outcome sequence:

See an object → predict an outcome → act → feel contact → observe success or failure → revise the policy

In that sense, Physical AI learns more like an animal or a human child. Intelligence develops not only by reading about the world, but by entering it through a body and learning from success, failure, and repeated correction.

Yet robots cannot be allowed to explore the real world without limits. A factory mistake can stop a production line. A household mistake can break an object. A mistake in medicine or eldercare can injure a person.

The practical path is therefore staged: learn first from human demonstrations and teleoperation; rehearse rare or dangerous situations in simulation and digital twins; validate in controlled physical environments; then collect failures, interventions, and outcomes from deployment. Only after review, retraining, and safety validation should new capabilities return to the field.

LLMs read the world humans have recorded. Physical AI must enter the world — and bear the consequences of every action it takes.

2. How Far Has It Progressed?Humanoid robots have moved beyond the stage of simply standing for a demonstration. Bipedal locomotion, balance, navigation inside factories, and repetitive material-handling tasks are approaching practical use in structured environments.

Unitree says it delivered more than 5,500 humanoid robots in 2025 and produced more than 6,500. UBTECH says its Walker S2 reached small-scale mass production and delivery at the thousand-unit level.

These figures matter, but production is not shipment; shipment is not end deployment; and deployment is not sustained productivity. Customers may be factories, but they may also be universities, laboratories, systems integrators, government demonstration programs, or entertainment buyers.

The model layer is advancing just as quickly. Google DeepMind’s Gemini Robotics can translate visual information and language instructions into action, with demonstrations including folding origami, packing food, and handling clothing. It can also replan when an object moves or slips.

But the strongest systems remain largely in research, private preview, or selected partner testing — not in unsupervised, long-duration household use.

The more mature capabilities today are locomotion, perception, and constrained tasks in structured settings. The harder problems remain dexterous hands, contact-rich manipulation, cross-environment generalisation, long-duration autonomy, and recovery after failure.

Models are beginning to learn how to perform a task. The industry still has to prove that it can repeat it every day, at an acceptable cost, without a human constantly rescuing them.

3. China Is Turning Factories Into Schools for RobotsChina’s most important advantage may not be lower component costs or faster manufacturing alone. It may be pursuing a different path from the search for a perfect general-purpose robot: enter structured production lines first, perform narrow tasks, and learn from deployment at scale.

A factory can rearrange racks, routes, lighting, and work heights around a robot. A task can be reduced to moving a specific container, sorting a fixed set of parts, or completing one repetitive assembly step. The machine does not have to understand every human environment on day one. It only needs to become more stable, safer, or less expensive in a defined setting.

UBTECH’s Walker S2 is designed to exchange its own batteries in about three minutes, enabling long-duration operation. But energy continuity is not the same as productive autonomy.

“Working for more than ten hours” can mean remaining powered on, continuing to move, repeating a fixed routine, or recovering autonomously when something goes wrong. What a customer should measure is closer to:

Effective autonomous hours = uptime × task success rate × unsupervised share × quality yield

Then subtract battery exchange, calibration, maintenance, recovery, and task-switching time.

What China may be accumulating is not just robot volume, but failure data, edge cases, and records of human intervention. The scarce data are often not how many times a robot succeeded, but why it failed, when a person took over, and how the system recovered.

Only when those experiences flow back into models and control systems can shipment volume become learning speed.

Factories are not merely a market. They may become the schools in which Physical AI learns to work.

4. Can Tesla Create a Second “Tesla Moment”?Tesla already has capabilities that matter: multi-camera perception, onboard inference, automated labelling, simulation, data pipelines, motors, batteries, power electronics, and large-scale manufacturing. It also owns factories that can serve as Optimus’s first customer and training ground.

Autonomous-driving technology can clearly help a robot see and move. Object detection, depth and velocity estimation, occlusion reasoning, occupancy modelling, motion prediction, and path planning all have transferable value.

But there is a fundamental difference:

Autonomous driving is largely about avoiding contact with the world. Robotics must learn how to touch the world safely — and change it.

A car mainly controls direction and speed. A robot must maintain whole-body balance while reaching, grasping, applying force, turning an object, and correcting itself from contact feedback. Driving video cannot substitute for data on fingertip slip, force, material properties, or failed grasps.

Tesla’s advantage is real, but “it has autonomy” does not mean “it has solved humanoid robotics.” Optimus still has to demonstrate effective labour cost, the share of time without human intervention, task-switching speed, joint and hand lifetime, and whether broader deployment improves the entire fleet.

It is also necessary to separate the roadmap from reality. Tesla’s 2025 annual report still described Bots as not yet commercialised. Its 2026 disclosures listed Optimus manufacturing facilities as under construction or preparation. Million-unit — and longer-term ten-million-unit — line designs are capacity ambitions, not current output, external delivery, or autonomous working hours.

A Tesla moment will not be established by a dance or a carefully edited video. It will be established by deployment data.

5. The Real Bottleneck Is the Hand, Not Only the BrainMusk is right to emphasise the human hand. Door handles, cups, scissors, tools, clothing, and appliances were designed around it. A machine with a truly capable human-like hand could use the existing human environment instead of forcing the world to be rebuilt around robots.

But human-like appearance is not human-like capability.

A useful robotic hand needs sufficient degrees of freedom, tactile sensing, force control, compliance, and durability. Degrees of freedom tell us whether a finger can reach a pose. They do not tell us whether the hand knows an object is slipping, whether it is squeezing too hard, or where contact is occurring.

A July 2026 project from Queen Mary University of London points to an intriguing solution. Its mechanochromic tactile sensor turns mechanical interaction into colour. When a soft surface is pressed, stretched, or sheared, its microstructure changes, producing visible optical patterns. A conventional camera can read those patterns as maps of contact, deformation, and pressure.

A robot’s camera may not only look outward. It can be placed inside the finger and use light to see touch.

The case also shows why Physical AI will not be delivered by bigger models alone. Materials, optics, sensors, mechanical design, and control must close the loop together.

Twenty-two degrees of freedom address whether the hand can move like ours. Optical touch addresses whether it can feel like ours. Without the second, the first may produce a more agile form of blind grasping.

6. Medicine and Care: Extend the Human Before Replacing OneRobotics discussions often begin with factory labour. In an ageing society, however, the more urgent shortage may not be a pair of hands for assembly, but an observer who can remain present.

The risks older people face are often not one dramatic diagnosis. They are small, continuous events: a missed or repeated dose, getting out of bed at night, a gradual slowdown in walking, reduced activity, or a fall after which no one hears the call for help.

The first value of Physical AI in care is therefore to remind, confirm, observe, record, escalate, and connect a person to a human caregiver.

Studies of in-home robotic medication systems suggest that machines can dispense scheduled medication packages, record missed retrievals, notify care providers, and reduce some routine home visits. But the boundary is important. Older adults are generally more comfortable with a robot reminding them to take medication than deciding which medication they should take. Reminder, dispensing, confirmation, and clinical judgment are not the same function.

Falls require the same discipline. The first realistic goal may not be a humanoid that physically lifts an older person from the floor. Lifting and transfer involve high-force, high-risk contact; one mistake could cause a second injury.

A more practical first step is detection followed by confirmation and escalation. The system recognises a fall or unusual inactivity, asks whether the person needs help, and alerts family, care staff, a nursing station, or emergency services if there is no response. A mobile robot could turn on lights, maintain a conversation, transmit only the necessary visual information, or bring a phone.

The 2025 TELEHPAD multicenter randomised trial enrolled 213 nursing-home residents with a mean age of about 87.7. The group receiving automated remote monitoring plus personalised secondary prevention had 0.24 serious falls per person-year, versus 0.49 in the control group. The share experiencing at least one serious fall was 20.19%, versus 33.03%.

This does not prove that a machine “prevented every fall.” It supports a more practical mechanism: detect events earlier, then let a care team initiate targeted prevention.

AI prediction itself remains immature. A 2026 systematic review of 28 longitudinal studies found a pooled AUC of about 0.79 for predicting future falls among community-dwelling older adults. Yet heterogeneity was extreme, only one model had undergone external validation, and every included model was judged at high risk of bias. Wearable gait data appear more useful as a complement to clinical and mobility assessment than as a replacement.

AI can turn daily change into a readable risk signal. Clinicians, therapists, caregivers, and older adults must still decide how to intervene. The real endpoint is not an AUC — it is whether falls, injuries, hospitalisations, and the loss of independent living are reduced.

Surgical robotics shows the same “extend before replace” trajectory. Sony’s microsurgery prototype scales a surgeon’s hand motion down to multi-jointed micro-instruments. SRT-H, a research system, autonomously performed the clipping and cutting steps of a gallbladder-removal procedure on ex vivo pig tissue and attempted to recover from errors such as missed grasps or instrument misalignment.

Neither case means that AI can independently operate on a patient. Sony’s system remains an unapproved prototype. SRT-H was not tested in human clinical surgery, in a living subject, or across a complete operation. Together, however, they outline a capability ladder: first amplify human action; then allow a machine to complete a bounded step; always preserve the ability for a person to intervene.

Good Physical AI does not take responsibility away from people. It extends human perception, reach, precision, and presence when our senses, time, or physical capabilities reach their limits.

The first breakthrough in robotic eldercare may not be lifting someone after a fall. It may be ensuring that a person no longer remains on the floor for hours without anyone knowing.

7. From Periodic Checkups to a Health TrajectoryModern medicine remains the irreplaceable centre of eldercare. But medical examinations occur at specific moments, while the condition of a person in her nineties may change gradually between visits.

Physical AI could connect those isolated medical moments into a daily health trajectory: walking speed, step length, left–right balance, time required to stand, nighttime bed exits, activity range, sleep, and eating patterns. With clinician direction and informed consent, it may also integrate blood pressure, heart rate, oxygen saturation, body weight, or glucose data.

The goal is not to turn an older person into a patient surrounded by dashboards or to let AI diagnose her. The useful objective is to establish a personal baseline and detect persistent change relative to her own normal condition.

A slower gait, more frequent nighttime rising, and a shrinking daytime activity radius may each remain below a clinical alert threshold. On a shared timeline, however, they may justify a closer look by family or clinicians.

Daily sensing → personal baseline → trend or anomaly → graded confirmation → human judgment → intervention

Without a confirmation method, an accountable person, and a response workflow, more health data will produce only alarm fatigue and anxiety.

Science is actively exploring whether ageing mechanisms can one day be slowed or partly reversed. David Sinclair’s work proposes that the loss of epigenetic information may be a major driver of ageing and investigates whether some of this change is reversible. It is an important research direction, but the claim that ageing is itself a treatable disease remains contested rather than a medical consensus. Laboratory reprogramming, animal lifespan results, or biomarker changes are not proof of a safe extension of human healthspan.

Perhaps future biotechnology will slow or reverse parts of biological ageing. Until then, Physical AI can help us notice functional decline earlier, reduce avoidable harm, and extend the period in which older people can live safely and with dignity.

8. Will People Buy a Worker — or a Companion?An industrial robot is valued through output, yield, availability, and return on investment. A household robot faces a harder environment: children, pets, clothing, glass, liquids, and temporary obstacles constantly change the task.

A companion humanoid may not need to solve all household work first. Its minimum viable value could be looking at a person, remembering her, responding to emotion, and creating a sense of presence through voice, expression, posture, and well-timed initiative.

AI on a computer waits to be opened. Embodied AI can remain in a shared space. It adds gaze, distance, approach, posture, and touch. Its defining capability may not be a benchmark score, but embodied presence.

The stronger that presence becomes, the stronger governance must become. For an older person experiencing loneliness, cognitive decline, or dependence on care, a system should not deliberately blur the distinction between a simulated relationship and a human one. It must not exploit attachment to drive spending, compliance, or consent to data use.

China is already testing demand. In June 2026, UBTECH introduced its U1 series of full-size lifelike humanoids at prices ranging from RMB119,800 to RMB990,000, and announced 13,361 cumulative orders. This is a market signal worth tracking, but at the time of writing it remains a company-announced preorder/order figure — not completed delivery or evidence of long-term use.

A deposit proves attention and initial willingness to buy. Stronger evidence will be full payment, delivery, three-to-twelve-month retention, subscription renewal, low returns, and support costs that do not overwhelm the business.

China may have shown that people will place a deposit on AI with a body. It has not yet been shown that they will choose to live with it over time.

Care also raises a subtler question: does a robot need to look more human to be more useful? For an older adult, a wheeled machine with a gentle voice, a clear screen, and simple arms may be safer, more stable, and easier to maintain than an expensive full-size bipedal humanoid.

The real test is not which form is most impressive. It is how much embodiment is needed to improve adherence, reassurance, interaction, and successful escalation when help is required.

9. What Must Never Be Outsourced to a MachineWhen Physical AI enters homes and care settings, it will collect more than text. It may observe how a person moves through a room, when she sleeps, what medication she takes, whom she speaks to, when her mood changes, and how her body is touched.

These data may improve care. They may also create unprecedented capacity for surveillance and manipulation.

At minimum, several boundaries should remain non-negotiable:

  • Older adults and families should know what is collected, who can see it, and how long it is retained.

  • Raw images should be processed on-device whenever possible, with only necessary alerts or authorised summaries transmitted.

  • Missed detections, false alarms, network outages, and device failures require explicit backup procedures.

  • A robot must not alter medication or replace qualified clinical judgment.

  • Emotional interaction must not exploit loneliness, cognitive decline, or dependency to induce purchases.

  • A user must always be able to stop the system, take it offline, or call a person.

  • Families must not mistake the presence of a device for the presence of a human who is actively watching.

Technology can extend human presence. It must not become an excuse for people to withdraw from a relationship.

Conclusion: We Need Systems That Respect Humans, Not Merely Machines That Resemble ThemChatGPT demonstrated that machines can generate human language. Physical AI must demonstrate that machines can understand human space, use human tools, and perform reliably without a person specifying every movement.

The era will not begin with the first robot that dances, folds a shirt, or imitates conversation. The real threshold will be crossed when a system simultaneously demonstrates effective autonomous hours, reliability, viable unit economics, and a learning curve that improves with deployment.

But if we measure Physical AI only through productivity, we may miss its most important value.

For an older person getting out of bed alone at night, what matters is not the robot’s valuation, degrees of freedom, or model ranking. What matters is whether it sees her lose balance; whether it can ask, “Are you all right?”; whether it knows whom to call when there is no answer; and whether, before the family arrives, it can tell her that someone is already on the way.

Robots may not first enter every home as workers. They may arrive as an interface that remembers to remind, refuses to ignore a call for help, and brings a human caregiver back into the situation. They may also become an embodied presence that remembers, looks back, and responds.

When AI leaves the screen, what we purchase may not be only a new form of labour. It may be a new form of care relationship.

What is worth hoping for is not simply that machines finally become more human. It is that technology may help us care for one another longer — and with greater dignity — when human capability reaches its limits.

Author NoteThis essay began with my family’s experience caring for my mother in her nineties. It is an editorial analysis of technical capability, deployment economics, medicine and care design, and social governance. It is not an endorsement of any company or product, and it does not predict a specific commercialisation timeline. Throughout the article, company statements, research prototypes, preorders or orders, capacity plans, actual delivery, long-term deployment, and independent validation are treated as different levels of evidence.

About the AuthorThe author writes about Physical AI, robotics, technology strategy, eldercare, health technology, and the social implications of emerging technologies.

DisclosureThis article was written independently and was not commissioned, sponsored, or reviewed by any company or organisation mentioned in the article. The author received no compensation from the companies discussed in connection with its preparation or publication and has no material financial, advisory, consulting, research-funding, or commercial relationship relevant to the subjects covered.

References to companies, products, researchers, and institutions are included for analysis and do not constitute endorsement. All opinions and interpretations are the author’s own.

Last Editorial ReviewLast editorial review: 11 July 2026

References1. Stanford HAI, What is Spatial Intelligence?

  1. World Labs, About

  2. NVIDIA, What is Physical AI?

  3. Google DeepMind, Gemini Robotics

  4. Tesla, 2025 Annual Report and 2026 company update

  5. Unitree, Clarification Regarding 2025 Sales Data

  6. UBTECH, 2025 Annual Report and Walker S2

  7. Sasso et al., High-Resolution Real-Time Mechanochromic Tactile Sensors, Science Advances (2026)

  8. Rantanen et al., In-home robotic medication management pilot study

  9. Sawadogo et al., TELEHPAD randomised controlled trial

  10. Gao et al., ML/DL models for predicting future falls, systematic review and meta-analysis

  11. Wu et al., Wearable gait features versus intrinsic fall-risk indicators

  12. Sony Group, Microsurgery Assistance Robot

  13. Kim et al., SRT-H, Science Robotics (2025)

  14. WHO, Healthy ageing and functional ability and clarification on old age in ICD

  15. Harvard Medical School, Sinclair Lab

  16. Xinhua, UBTECH U1 launch pricing and company-announced orders

Further Reading- Spatial intelligence and three-dimensional world understanding: Stanford HAI, What is Spatial Intelligence?

DisclaimerThis article is for general informational and educational purposes only. It is not medical, clinical, investment, legal, engineering-safety, or emergency-response advice. Company figures may be self-reported, based on preorders or announced orders, or presented as forward-looking capacity plans; product availability, performance, regulatory status, and research evidence may change. Decisions involving eldercare, fall detection, medication management, surgery, physical contact, or emergency escalation should be evaluated by qualified professionals and supported by appropriate human oversight, fail-safe procedures, and clear accountability. In an emergency, contact the local emergency services for your location.

HashtagsCore: #PhysicalAI #Robotics #EmbodiedAI #Eldercare #HumanInTheLoop

Extended: #SpatialIntelligence #HumanoidRobots #DexterousManipulation #HealthcareAI #CareTech #AIForGood #TechnologyStrategy #FutureOfWork #Tesla #ChinaTech

This article was originally published on Medium. Read the full version with charts and figures →