What it is
A policy is the executable decision layer. It may be end-to-end, modular, or hybrid with planners and controllers.
Why it matters
Most Physical AI claims eventually have to pass through a policy that moves hardware.
How not to overread it
Policy behavior is embodiment- and environment-dependent.
Related edges
Vision-language-action models
Robot action generation
Policy behavior is embodiment-specific.
Cross-embodiment data
Cross-robot learning
Transfer claims need per-embodiment evidence.
LeRobot
Policy training
Reproducibility still depends on hardware and data.
State estimation
Policy inputs
Bad state estimates can make good policies fail.
Manipulation
Object interaction
Narrow task success is not broad reliability.
Control stack
Hardware commands
Model output must respect control constraints.
Onboard compute
Onboard inference
Compute spec does not imply task success.