t5lab.planner_arms¶
Phase-2 planner ARMS — alternative L2 value-field planners that match the MVProp interface
(.field(x) -> (H,W), .move_scores(x, rc) -> (5,)) so they drop straight into the same
distillation trainer, generalization eval, and MVPropPlanner wrapper. One-file zoo + factory.
Arms (why each — the ablation question):
* HighwayMVProp — MVProp + a per-cell HIGHWAY carry gate (Highway Value Iteration Networks,
Wang et al. ICML 2024). V_{k+1} = g·max(r, γ·w·max4(V_k)) + (1-g)·V_k. The (1-g)·V_k
skip is a GRADIENT HIGHWAY across the K sweeps, the published fix for the vanishing gradient
that kills deep value iteration. THE mandatory control: does the architectural fix remove the
need to distill (i.e. does RL-only now train)?
* GPPN — Gated Path Planning Networks (Lee et al. NeurIPS 2018). Replaces VIN's hard max with a
DATA-DEPENDENT (LSTM/GRU-style) update gate computed each sweep from [V_k, candidate,
neighbour-mass], for stable long-horizon credit propagation.
* MSP — a compact Spatial-planning transformer: AXIAL self-attention over grid rows/cols (size-
invariant, O(HW·(H+W)) not O((HW)²)) with sinusoidal 2-D positional encoding predicts the
value field DIRECTLY — no explicit value-iteration recurrence. Tests whether attention can
learn to route.
All share MVProp's input x = (2,H,W) = [blocked, goal_onehot] and are fully convolutional /
axial → SIZE-INVARIANT (same weights at any grid), so they compare fairly with MVProp.