cleaningPolicy

LEGACY BASELINE ONLY: no longer imported by the live Roomba. Current control is flyNeuralPilot with local sensing and trained GPU synapses. Historical Q-table tests below do not describe the current pilot.

Goal-conditioned tabular Q-learning over 466 collision-safe states and 72 mess goals. 180,000 exploratory transitions with random resets, gamma .97 and alpha .35; observed transitions are replayed for each goal. No pathfinding or scripted route. Policy is trained for this fixed house and dirt layout, and continues TD updates on reached live states. Evaluation separates trained versus random actions and requires three visits for fluid sites. Use train, report, choose and mask.