PASC.

Physical AI Safety Consortium
A coalition of universities and industry building one shared definition, one benchmark, one standard for the safety of AI that acts in the physical world.
Become a founding member →
A sentence can be retracted. A motion cannot.

The Six Problems of Physical AI Safety

Physical AI safety asks whether AI remains governable once its decisions enter a closed loop with a body and the world. Model safety is necessary. Once the model can act, it is not enough.

I.

Specification

Whether harm can be defined precisely enough to enforce, without stripping away the context that makes it harmful.

II.

Anticipation

Whether an unsafe consequence can be predicted, with calibrated confidence, before the last moment it can still be prevented.

III.

Falsification

Every test is finite. The world is not. Whether failures beyond the test can be found, and what passing can ever prove.

IV.

Adversary

Whether any safety guarantee survives an opponent inside the loop.

V.

Recoverability

Which states still admit a way back, and whether the boundary can be found before the machine crosses it.

VI.

Invariance

Which guarantees, if any, survive a change of body, of task, of world, of scale.

What we build 01

Physical AI safety is a stack. Every layer is unfinished.

Define
In preparation

The definition

A position paper that states the six problems formally and defines safety as a state to be maintained, not a verdict to be issued. Problem I.

Measure
In design

The benchmark

Standardized evaluations of anticipation and recoverability, scored across embodiments. Problems II, V, and VI.

Standardize
Planned

The standard

Runtime monitors and reference implementations that make the benchmark a bar to clear, not a badge to claim. Problems III and IV.

Who's involved 02
Founding member organizations
Publications & Projects 03
Publications
Projects
Problem I

The definition

A formal account of physical AI safety as a state to be maintained.

Problem II Problem V Problem VI

The benchmark

Evaluations of anticipation and recoverability across embodiments.

Problem III Problem IV

The standard

Runtime monitors and reference implementations that turn evidence into a bar.

Six problems stand between physical AI and the world. Pick yours.

Become a founding member →