Physical
AI.

I. Intelligence acquires consequences

To act is to change the problem.

An agent answering a question can often retry. A robot moving through a crowded room changes the world while it reasons about it. Its sensors are incomplete, objects move, and its actions have consequences that cannot always be undone.

Embodied systems combine estimation, planning and control under tight time constraints. The most interesting scaling question is not how to give one robot more intelligence, but what becomes possible when many embodied systems—and people—must share a physical world.

II. The boundary is not a packet

Delay changes correctness, not just performance.

Consider two robots handing a part from one to another. If one acts on an outdated estimate of the other's pose or motion, an otherwise correct plan may fail. The safe timing budget depends on the speed, geometry, uncertainty and control design; there is no universal millisecond threshold.

A distributed physical system needs more than transport: a useful account of position and orientation, consistent reference frames, synchronized clocks, uncertainty estimates, action intent and the ability to detect when shared state is no longer trustworthy. Local autonomy remains crucial, especially when links are slow or unavailable.

III. From fleets to collectives

Many robots are not automatically one system.

A fleet manager can dispatch machines without enabling them to jointly perceive, negotiate or adjust motion. More demanding settings—industrial handoffs, shared workspaces, logistics facilities or disaster response—require participants to coordinate dynamically with other machines and humans.

The hard problems occur at the seams: which participant has the right of way, how intent is communicated, who may override a maneuver, and how the group reacts when observations disagree. Interoperability and common representations help, but coordination also depends on protocols for trust, safety and accountability.

IV. An unfinished world model

What must everyone know—and when?

Perfectly shared, perfectly current world state is unattainable. Sensors disagree, communication delays vary, and a centralized model may become stale before an action completes. A practical system must decide what to share, at what precision, with which guarantees, and when to fall back to independent safe behavior.

Work on multi-robot localization, distributed mapping, formation control and shared autonomy provides pieces of the puzzle. Real deployments remain constrained by robustness, heterogeneous hardware and safety assurance. The frontier is the architecture that can compose these pieces across organizations and environments.

V. Toward physical societies

Can intelligent motion become a collective capability?

As machines move among people, coordination becomes social as well as spatial. They must interpret human intent without assuming it is always explicit, respect norms and constraints, and remain accountable to human authorities.

The possibility is larger than coordinated automation: populations of people and embodied agents that learn how to work and move together. That is a research direction, not a demonstrated general capability. Its promise must be judged by safer, more reliable collective outcomes—not the number of connected robots.

A moving system cannot afford to share only yesterday's truth. Intelligence must stay aligned in space and time.