Occupancy A/B · 32×32 / 10 agents · 2 arms × 4 maps × 2 seeds (16 runs)

Occupancy A/B — does a richer belief help, or hurt?

What it tests: a head-to-head between two perception arms on all four maps, two seeds each. SLAM-only (the baseline) senses only walls and gossips them. occupancy adds a full occupancy belief — sense_free (mark cells known-empty), a boundary channel, and an occupancy-frontier signal (Yamauchi-style) — strictly more information per step. The question is whether feeding the policy that richer map actually buys coverage.

Key finding Measured live from the rendered tiles below. Occupancy is worse, not better. On all four terrains and both seeds the richer occupancy belief loses to plain SLAM-only by ≈ 3–7 points of coverage. A clear negative lever: the extra channels do not pay for themselves here.
Per map — mean of 2 seeds · rollout coverage = final % of free cells visited (100-step re-roll) · Δ = occupancy − SLAM-only
MapSLAM-onlyoccupancyΔ (occ − base)seeds
computing from the rendered rollouts…

The run list is grouped by arm, then map — expand SLAM-only vs occupancy to compare the same map under each belief. The table is computed live from the rendered rollouts, so it self-corrects on every re-render.

t = 0