Physical AI, 2035

Between now and 2035, the centre of gravity in intelligence will move from tokens to contact. Not because language will become uninteresting, but because the remaining work of civilization is physical: care, construction, logistics, energy, and the machines that fail when a model is merely eloquent.

Physical AI is the name we give to systems that can perceive, predict, and act under gravity. It is not a humanoid. It is not a warehouse arm. It is a layer: world models, foundation control, closed-loop data, and assurance that can survive a lawyer and a physicist in the same room.

The curriculum of reality

Text is a lossy compression of the world. Useful, cheap, infinitely copyable. Reality is expensive and unforgiving. The laboratories that will matter are factories, hospitals, farms, and circuits — places where a wrong action has a mass. Motorsport is one such laboratory. So is a ward at 3 a.m. So is a line that cannot stop.

A system that cannot fail safely has not yet earned the right to move.

What 2035 requires

• Generalist control that transfers across embodiments without becoming a stunt.
• Simulation that knows it is lying, and a data loop that corrects it.
• Permission systems shared with digital agents, stricter by an order of magnitude.
• Public evaluation: not a video, a protocol.

Celestra’s physical program is paced to that decade. We will not hurry a body ahead of a brain that can be responsible for it. We will also not pretend that intelligence which cannot touch the world is finished.