A machine self-model predicts how a bounded machine will respond to a candidate command. It gives a controller, planner or operator a way to compare consequences before acting and to notice when the machine no longer behaves as expected.
State, action and response
Telemetry describes what is happening now. A self-model adds the missing relationship between the current state, an available action and the state likely to follow. On a legged robot, that can include joint position, velocity, contact and body motion conditioned on a control command.
Action conditioning is essential. A system that only labels the current state may be useful for monitoring, but it cannot compare what different commands are likely to cause.
The model is bounded
A useful self-model names the machine, signals, controls, time horizon and operating conditions it supports. It should also expose uncertainty and refuse inputs outside that boundary.
This makes the claim testable. The model is not a general understanding of physics. It is a predictive representation of one machine or compatible machine family under declared conditions.
Separate from the controller
The controller decides which action to execute. The self-model predicts what candidate actions are likely to do. Keeping those responsibilities separate makes it possible to improve or repair the predictive layer without silently replacing the robot's control and safety stack.
A08 is built around this boundary. It maintains the self-model and its evidence. Execution authority stays with the existing system.
Sources
World Models ↗A foundational paper on learned environment models for prediction and control.
A Path Towards Autonomous Machine Intelligence ↗A predictive architecture built around world models, objectives and action planning.