Skip to content

The balance-beam sweep

Question

Does the structure of the modulatory coupling and the body geometry determine the structure of the world the agent can build?

Setup

Each condition runs the self-localized mapper under a different "balance beam", varying both:

  • (a) information-routing topology — how whisker affordances are routed into the steering command: flat/distributed (all whiskers pooled at once) vs hierarchical/layered (whisker groups summarized, then combined), via HAPExplorer(routing=...);
  • (b) physical body geometry — whisker count (3 / 5 / 9) and drive-zone track width (narrow / wide), via MouseSchema;

plus the architectural toggles ±BAP (the locomotor drive) and ±HAP (the affordance-recruited haltable action), and proprioceptive noise.

Conditions (9)

full · hierarchical routing · sparse (3w) · dense (9w) · narrow track · wide track · −BAP · −HAP · noisy proprioception.

Run

cd experiments && ../.venv/bin/python p2_topology_sweep.py

Outputs

  • figures/p2_topology_sweep.png — a 3×3 grid of the constructed maps.
  • data/sweep_results.csv — coverage, precision, drift, and parameters per condition; per-condition data/sweep_*.npz (hit clouds + trajectories). The full data is gitignored; a curated copy is in samples/.

Findings

Balance-beam sweep — a 3×3 grid of constructed maps across routing × morphology × ±BAP/±HAP × proprioceptive noise

A 3×3 grid of constructed maps across the swept conditions. The columns and rows that build a faithful map are immediately visible; the ones that collapse it (no BAP, no HAP, heavy proprioceptive noise) are equally visible.

  • ±BAP / ±HAP are decisive. Removing the basal drive (−BAP) collapses coverage to ~20% (no locomotion → it maps only a sliver); removing the haltable affordance-action (−HAP) drops it to ~44% (drives into a wall and stalls). Both modulatory layers are necessary to build a world model — "no modulation → no learning."
  • Self-localization quality gates fidelity. Noisy proprioception keeps coverage but drops precision (100% → 74%) with ~6.7 cm drift: the snapshot smears.
  • Routing and morphology are robust here (~96–97% across flat/hierarchical, 3/5/9 whiskers, narrow/wide track) — the arena is easy enough that the architecture is forgiving. Harder tasks (sparser sensing, a maze, faster motion) would separate these axes; that knob is built in.