Humanoids in Australia, chapter 1. v1 preliminary, 2026-08-31. How the numbers are made is on the method page.
The Long-term Vision
A humanoid that reminds your elderly mother to take her tablets, notices when she falls, helps her up and calls a nurse: we expect supervised pilots in Australian aged care homes around 2030. A humanoid resides in the family home, that gets her out of bed and walks her to the bathroom: probably not this decade. Our central estimate for the first supervised pilots is 2039, and in 98% of our simulation runs nothing like that is in commercial use by 2045.
That is the short version. The rest of this page is why, and what would change our minds.
The robot is not the problem
The cheap humanoids are real. Unitree sells one for about US$13,500. AgiBot is launching in Australia at US$40,000 to $50,000 a unit. 1X is taking orders for a home robot at US$20,000, or $499 a month. Unit costs are falling somewhere between 20 and 40 percent a year depending on whose numbers you use, and Bank of America expects a Chinese-built humanoid to cost under US$17,000 by 2030.
None of that gets a robot into an aged care home. To do the medication job it needs a reliability record, a privacy case, a funding line and a nurse willing to let it do the round. To do the mobility job it needs all of that plus a safety standard for a machine that touches a frail person. There isn’t one. Not in Australia, not anywhere.
COUNTER: You’re talking like government will be able to stop it from happening. What if this plays out like Uber? Humanoids WILL be in the home. The AI will be powerful. People will GIVE the humanoids the authority to help out, at first around the house, doing chores, cooking, cleaning, pet minding, gardening, house repairs, home security, etc. That’s where they will be tested. That’s where trust will build. The question is how long before people allow them to mind the kids while they are out. At first, for 30 minutes. Then an hour. Then a day. And then how long before they allow them to mind their elderly parents. Ten minutes. Thirty. A few hours. And then permanently.
The international standard for personal care robots, ISO 13482, has never been adopted as an Australian standard. The new standard being written for humanoids, ISO 25785-1, covers factories only and says so in its scope. Nobody has certified a legged humanoid to either. No vendor publishes reliability data. No Australian insurer offers a product. No funding scheme lists a robot. And nobody has written down who is accountable when a robot performs a transfer at three in the morning and it goes wrong.
Any one of those stops the job.
COUNTER: You’re assuming people don’t just buy them. Who is going to stop them from giving them power over their kids and parents? Can a government even do that? Realistically?
That is the whole finding, really. Adoption is a chain, and the chain moves at the speed of its slowest link.
COUNTER: No, it moves at its FASTEST link. Early adopters do it. Then they tell their friends. It’s the adoption curve.
Three jobs, three timelines
2026202920322035203820412044Medication reminders and verificationMobility assistance53% never98% neverManipulation for ADLs94% neverpilot grade, P10–P90commercial grademedian (P50)
Each bar is the range of years, across 1,500 runs of the model under four scenarios, in which that job first reaches pilot grade. The mark is the median. The darker bar is the same thing for commercial grade. Where a bar runs off the right edge, most runs never got there by 2045.
Medication reminders and monitoring is the near one. The talking, recognising and reminding parts are done. Andromeda’s Abi is in dozens of Australian facilities doing exactly this on wheels, and Korea has twelve thousand Hyodol dolls doing it in pensioners’ homes. What is missing for a humanoid version is a reliability record, a settled privacy position, and a funding door. Pilot grade around 2030 (the band runs 2028 to 2032), commercial use around 2038.
Fetching, tidying, helping with meals and dressing is the middle one. The fetch-and-carry half is close. The touching half is not. The best published dexterity numbers, from DeepMind in July 2026, are 92 percent for unscrewing a light bulb and 40 to 45 percent for sealing a bag or picking something off the floor. Dressing research is still mostly done in simulation. Pilots around 2037, band 2034 to 2043.
Transfers, walking support and fall response is the far one. No humanoid anywhere has demonstrated a bed-to-chair transfer of a person with published results. Japan’s RIKEN built a bear-shaped robot to do it in 2015 and retired it as impractical. Pilots around 2039, band 2035 to 2044.
Japan already ran this experiment
Japan has spent twenty years and more than US$300 million subsidising care robots. About three quarters of its nursing homes now have a monitoring sensor. A quarter have some kind of mobility or transfer aid. Almost none have a robot that does hands-on care. And the robots that were adopted did not replace anyone: a Stanford study published in August 2026 found homes that adopted them employed 28 percent more care workers and 39 percent more nurses, and had fewer staff quit.
That is the shape we expect here, a decade later and with better robots. The near-term story in Australian aged care is reminders, monitoring and fetch-and-carry, and it starts around 2030. The hands-on story is a 2040s story unless several things change at once.
What would move it
The obvious lever is not the one that works. Doubling the speed of regulation, which is roughly what an industry-union accord would do, brings the medication timeline forward by about three years and the physical tasks by about one. An accord cannot fix the fact that no robot can safely take a frail person’s weight yet. What it can do is decide who writes the safety standard, the supervision model and the evidence base, and whether Australia does that first or waits for someone else’s.
The other lever is boring and large. The median year in our simulation trails the best-guess path by about four years, purely because so many things are uncertain. Anything that reduces the uncertainty, starting with a vendor publishing a reliability record and a provider running a sanctioned pilot, moves the timeline as much as anything that improves the robots.
What would change our mind
We have committed in advance to what each of these is worth. When one resolves, the model gets the update written next to it, whether or not it flatters this page.
What this is
Version one, dated 2026-08-31. The numbers come from a model of 39 prerequisites across six areas, each scored from published evidence as at that date, combined so that the weakest ones set the pace, and run forward 1,500 times with every score and every speed drawn from its uncertainty range. The scores are starting estimates, not measurements. Every one of them is shown with its reasoning and sources in the working, along with the parts of the model we would push back on ourselves.
If you think a number is wrong, you can find it and say which one and why. That is the deal.
See the working → Every prerequisite, every score, every source, every speed, and the reasoning behind them. Members only.