Skip to content

Concepts: the SMN → simulation mapping

The bench implements components of the SMN architecture as controllers driving a MuJoCo body. The vocabulary below is used throughout the experiment pages.

Architecture → simulation

SMN component In the bench
Zone / CAZ (Coordinated Action Zone) a body segment + actuated DOF that senses and acts (dual-port).
S — transducer a rangefinder "whisker" (distance sensor); contact/gradient sensors are possible.
M — modulator a pull-only actuator; zones are opponent pairs (agonist/antagonist), giving a signed net activation.
N — communication board = the "balance beam" the coupling that routes sensed affordances into the drive/steering command — the experimental independent variable (HAPExplorer.routing, drive mapping).
BAP — basal action pattern a central pattern generator (CPG) providing the baseline locomotor drive.
HAP — haltable action pattern HAPExplorer: action recruited/halted by sensed affordances.
snapshot / world model the state the agent builds from (action, modulated sensation) — here an occupancy map.
reafference a forward model keyed on the agent's own state; its residual separates self-caused from world-caused sensory change.

Key ideas

  • Opponency. Each zone is an antagonist pair. This is not decoration: a body with only forward-pulling drives cannot rotate in place — opponent drive zones can (one forward, one back = pure torque). Steering and turning require it.
  • Body-relative world model. The agent knows where its zones are (the body schema) and localizes from its own proprioception; the world model is built in the agent's own frame, never from an absolute "god's-eye" pose. (See the P1 → P2 shift.)
  • The balance beam as the variable. The experiment fixes the body, world, and task, and varies the modulatory coupling + body geometry, asking: does the structure of the balance beam determine the structure of the world the agent can build?

The experimental logic

  • Control (fixed): body, world, task ("explore and build a model").
  • Independent variable: the balance beam — routing topology (flat vs hierarchical), body morphology (zone count/placement), the ±BAP/±HAP toggles, proprioceptive noise.
  • Dependent variable: the constructed world model — coverage, precision, and dead-reckoning drift.

For exactly what is simulated vs computed vs assumed, see Assumptions.