lewm-table01-3k

LeWM world model fine-tuned on Push-T with a damped ("table") board.

initialised from	c03
training corpus	desertmouse/pusht-simulator-table-mu01
physics	table_mu = 0.1 → per-tick damping (1-mu)^10 = 0.3487
schedule	10 epochs, lr 2.5e-5, 100 episodes held out
validate prediction loss	0.00000 → 0.00000

table_mu is not a Coulomb friction coefficient. The board applies exponential velocity damping to the block each tick, plus stop thresholds (2 px/s, 0.02 rad/s). table_mu = 1 means zero damping, i.e. stock-equivalent physics; smaller values make the block glide further. This mechanism was identified by replaying recorded actions and comparing trajectories, not assumed: at mu 0.3 replay matches to 0.2 px, against 11.3 px for stock and 16.9 px for a Coulomb model.

Contents

• weights.pt — final weights (72 MB, 18.04 M parameters)
• action_stats.json — the action normaliser this model was trained under
• config.json — training configuration
• train.log — full training log

Using it

Load weights.pt with this repo's own action_stats.json. Do not reuse a normaliser from another checkpoint: the scales differ between corpora, and a mismatched normaliser silently rescales every action the planner proposes.

Caveat

Lower prediction loss did not imply better control anywhere in this project. Evaluate planning performance directly before treating this checkpoint as an improvement over its parent.