Week of 22 August 2026
Physical AI and the environments it reaches first
A machine you do not own can still change the corridors, lifts and service flows you are responsible for
Physical AI can matter before an organisation buys a robot
Physical AI is AI that senses and acts in the real world. Unitree's humanoid demonstrations have made the category easy to see, and the more useful question sits behind the spectacle: when could a system that moves through, inspects or serves a real environment change a business you operate, depend on or govern? Property, hospitality and retail meet it first because they manage corridors, lifts, queues, rooms and sites. Others meet it through assets, suppliers, clients, insurance or a service partner.
The “Superman” preview: a 2 m standing jump and a top speed of 12.66 m/s.
It has no hands and no grippers. It sets records for jumping and running and cannot carry a tray, open a door or pick anything up.
The category is wider than a humanoid
A human-shaped machine is one option among several. Each of these already works where the environment is controlled enough, and each depends on something specific staying true.
The right shape follows the task, the environment, the safety requirement and the cost of adaptation. That is why a company can be exposed to this category through a supplier or a tenant long before it buys anything.
Two more demonstrations make the rest of the argument visible
Moving is the part that already works. These two are about the harder half: giving a machine a usable sense of the space it is in, and training a task before it has to survive one.
World Labs · Nov 2025
Marble
Turns text, images and video into a persistent 3D world you can move through and edit. This is the machine being given a space to reason about rather than a picture to describe.
World Labs notes that a world can look coherent and still lack the precision robotics needs.
World Labs · Jul 2026
Real-to-sim-to-real
One physical task rebuilt in simulation, varied, trained against, then run again on the hardware. This is how a task gets practised before it meets a real corridor.
Real and simulated runs side by side, at 2x speed.
A polished demo does not yet show a dependable service
Each of those clips is real, and each was recorded under conditions the vendor chose. The commercial question is what remains when the conditions stop being chosen.
A demonstration proves an action is possible. A service is what survives the two-hundredth run, and the distance between the two is where the cost sits.
What the clips establish
The action is possible. A body that jumps and runs, a space a machine can reason about, a task trained in simulation and run for real.
The hardware is real. Balance, grip, sensing and movement are capabilities you can watch rather than claims to assess.
The direction is credible. Enough progress in ten months to justify treating the category as a live question.
What a service still needs
Recovery. The machine notices the task is going wrong and either reaches a safe state or hands to a person who can fix it.
Maintenance and uptime. Someone services it, and the round still runs on the days it is out.
Integration. Lifts, doors, rosters and ticketing treat it as part of the operation rather than a visitor.
A named owner. A person accountable for the outcome, with the authority to stop it.
Why recovery is the hard part
Robot training data can hold the whole sequence in which a person fumbles an object, catches it and carries on. Train on a clean edit of a successful attempt and the machine learns what success looks like. Train on the unedited attempt and it learns what to do when the object slips. That second behaviour is the one a corridor, a lift lobby or a shop floor will ask for, and it is the one a demonstration reel is edited to remove.
Which is where world models come in
A world model is a representation of a place that a machine uses to work out what its next action would do there, updated as the space changes. A trolley appears, a guest turns into the route, a lift runs late. A machine with a usable one can tell that the corridor no longer matches the route it planned. World Labs, whose Marble clip you just watched, makes the limit explicit: a world can look coherent and still lack the precision robotics needs.
What this means for a leader
Judge a proposal on the conditions the vendor did not choose. Ask what the machine does when it is wrong, who it hands to, what it costs to keep running, and who is accountable for the outcome. Progress here arrives task by task, so a repeated job in a back corridor can be viable while the same machine on a customer floor is not.
Managed environments may shape where physical AI first works at scale
A building operator may never own a robot and still decide whether one can do useful work on the site. Six decisions set the terms, and each one shows up directly in what the service costs to run.
Routes
Which corridors, service areas and floors are open to it, and at what times.
A route agreed for the night shift and used at midday puts a machine into the busiest hour it was never scoped for.
Access
Doors, barriers and secure areas, and the credential the machine presents.
If a machine cannot open a door on its own, a member of staff is now part of every run, and the labour saving goes with it.
Lifts
Whether it can call a lift itself or waits for someone to press the button.
Manual lift access is usually the difference between a delivery robot that pays for itself and one that does not.
Charging
Floor space, power and the time it spends off the round.
Charging points sited for the building rather than the route add dead travel to every shift.
Hand-off
The point where the machine stops and a person continues the task.
An undefined hand-off shows up as a guest standing beside a stopped machine with nobody to ask.
Incident response
Who attends, who has stop authority, and what evidence is kept.
The answer is agreed before an incident or improvised during one, and improvisation is what reaches the customer.
Commercial buildings are likely to hold a mixed population of these machines, owned by tenants, contractors and service partners rather than by the operator. Whether that makes an operator a platform, and who captures the value, is genuinely open: privacy, tenant consent, liability and bargaining power will settle it site by site.
Board attention should follow operating triggers, not headlines
A board does not need a robot strategy. It needs to recognise the four moments when this stops being someone else's topic, and to know what evidence would persuade it that a demonstration has matured into a service.
A tenant asks for autonomous delivery access
A business inside your site wants machines moving through it. The access rules you set decide what service that machine can offer, and who answers when it fails.
A service partner arrives with an inspection device
A contractor brings a quadruped or a drone. Access, data boundaries and the response after an incident are settled before it walks in, or they are settled during the incident.
A client's economics change through physical automation
Insurers, lenders and professional-services firms meet this through a client's operating environment long before it appears as their own procurement decision.
A bounded task starts to move cost or capacity
Reliable performance on one repeated task changes what a site can promise and what it costs to promise it. That is the point the category becomes an operating-model question.
What leaders should take away
Physical AI will commercialise unevenly, and the first decisions it asks of an executive are about environments and responsibility rather than procurement.