t5lab.planner_generalization¶
Distillation-bias check: does the distilled MVProp planner generalize to map types it never saw in training? Training used random rectangles + open (gen_wall). Here we test STRUCTURALLY DIFFERENT maps — rooms, serpentine mazes, pillar lattices, DIAGONAL barriers (the hardest for a conv trained on axis-aligned rectangles), dense clutter, and the held-out SAR set if present.
Two metrics per map type, vs the 4-connected geodesic oracle: * optimal-move : fraction of (map, goal, cell) where the planner's argmax-V move achieves the minimum 4-connected distance-to-goal among the 5 neighbours (local correctness). * reach : fraction of (map, goal, start) where FOLLOWING the field from start actually arrives at the goal within a step budget (GLOBAL correctness — catches a field whose only flaw is a local maximum that traps the agent). This is the strong test.
Run: PYTHONPATH=. GRID=32 PLANNER=mvprop_distilled.eqx $PY -m t5lab.planner_generalization