ctde_v0.run_warmbig¶
Warm-start into a BIGGER world — does the occupancy belief make small→large transfer better? (User hypothesis: occupancy/boundary should pay off most when you scale up, because the policy learns general "chase the occupancy frontier, stay inside the field" rules that don't depend on grid size.)
Phase 1 trains @16²/4; phase 2 warm-starts @32²/10 from the MATCHING phase-1 checkpoint
(--init-from; the LPAC backbone is scale-invariant, so a 16² policy loads into a 32² run).
Two arms, held consistent across phases so the obs-channel count matches on load:
base : SLAM-only (5 obs channels)
occ : + --sense-free --boundary (7 obs channels: occupancy + occ_frontier + boundary)
Read against the from-scratch 32² numbers from run_occ — the question is whether the occ
warm-start climbs higher / faster (% of the per-map optimal) than the base warm-start.
cover_r=0, 100-step horizon, SLAM on, no hard mask.
PYTHONPATH=.:../../../FiedlerValueEstimation ~/ZymeraLab/.venv/bin/python ctde_v0/run_warmbig.py --out runs/warmbig --seeds 2 --jobs 1