ctde_v0.run_occ¶
Occupancy A/B — does the occupancy belief beat the SLAM-only baseline, per terrain?
Same fixed best recipe both arms (role explorer/relay split + learned-Lagrangian SOFT
connectivity + frontier-attn + SLAM wall-sensing, NO hard mask); the ONLY difference is
the belief the policy sees:
base : sense_walls only — walls in the belief, coverage-trail frontier (5 obs ch)
occ : + --sense-free --boundary — the FULL sensed region enters the belief (a real
free/occupied/unknown occupancy map), a true Yamauchi occ_frontier to chase,
and the field-edge boundary channel (7 obs ch)
@32²/10, cover_r=0 (REAL visited coverage, scored vs the per-map oracle optimal: open 72 / rooms 37 / mixed 31 / crowded 40), 100-step horizon. Reports coverage_pct + connectivity_real per (arm, terrain) over N seeds — so occupancy's lift (or null) is clean per map. The corridor maps (rooms/mixed/crowded) are where walls + boundary should matter.
PYTHONPATH=.:../../../FiedlerValueEstimation ~/ZymeraLab/.venv/bin/python ctde_v0/run_occ.py --out runs/occ --seeds 2 --iters 2000 --jobs 1