Skip to content

P1 — a multi-CAZ agent builds a world model

What it shows

A mobile "mouse" with a steering CAZ and a whisker fan explores an arena and constructs a map of it ("the picture") from its pose and whisker readings.

Setup

A planar agent (slide_x, slide_y, yaw) with a five-whisker rangefinder fan. Steering is an opponent yaw pair (two MuJoCo motors); forward locomotion is a central body-frame thrust applied via xfrc. A CPG (BAP) provides the drive; a HAPExplorer (HAP) steers toward open space and rotates when blocked. Each whisker hit is placed in world coordinates and accumulated in an OccupancyMap.

Assumptions specific to P1

(in addition to the common assumptions) - The map is built from the simulator's true pose (god's-eye). This is the first cut; it is not the agent's own self-knowledge. See P2, which removes this assumption. - Locomotion is a central thrust (not located drive zones). - The occupancy map is a monotonic accumulator (no decay).

Run

cd experiments && ../.venv/bin/python p1_world_model.py

Outputs

  • figures/p1_world_model.png — true arena outline, exploration path, constructed point cloud, and coverage/precision.

Result & interpretation

P1 world model — exploration path, constructed hit cloud, and the coverage/precision summary

Arena outline, exploration trajectory, the constructed hit cloud against the true surfaces, and the coverage/precision numbers. The walls and both objects are recovered at ~97% coverage and ~99% precision.

The agent reconstructs the walls and both objects at ~97% coverage, ~99% precision. Because the map uses the true pose, the reconstruction is essentially the upper bound — which is exactly why P2 rebuilds it from the agent's own self-localization, so the constructed world model is genuinely the agent's construction (and so the sweep is meaningful).