messSimulation

Conservative finite-volume wet floor model. Wheel contact transfers film into track streaks; suction alone removes mass. Shares its numerical model with sensor sampling and training. No fixed visit counter controls cleaning.

GPU wet-film visual reconstruction over a persistent mesh. Smooths the conservative floorFilm density, relaxes surface height with viscous lag and computes normals by central differences. Adapted from ZeroMind cloth_vertices by maker, MIT, 08a2b7d0-361c-40c2-82a0-1ac00a418a5d. No CPU mesh replacement during animation.

Reproducible three-room floor plan, obstacle collision queries and 72 seeded irregular mess locations. Nine fluid spills; remaining sites contain individual grit, paper or crumbs. Shared by learning and live simulation.

Glossy white PBR base with vertex colour supplied by the shared viscous floor-film compute surface. Coffee, paint and detergent colours blend by local mass; clearcoat highlights reveal deformed edges.

A tiny pilot. A whole home. Three furnished cutaway rooms with a fly-driven robotic vacuum, live neural render texture, pilot camera, individual debris and deforming coffee, paint and detergent spills. Uses the published fruitFly package, path_grit material and set_floorboards material. Learning is an added goal-conditioned Q controller, trained with 180,000 exploratory transitions in this fixed house. The biological connectome weights stay fixed. Desired wheel commands stimulate left/right motor neurons; measured GPU spiking rates pass through a calibrated decoder into differential-drive motion. No authored cleaning route or pathfinder is used. Scene coordinates and remaining mess are available to the controller; this is not a vision-only or sim-to-real robot system. Controls: Pause, New dust (reproducible layout), Silence/Restore motor neurons, Train again, Untrained baseline, room preference, Apartment and Follow fly + joystick. Grid validation: 72/72 sites from each of three starts, including three visits per spill; random-action baselines 24, 45, 46 of 72 within 1,800 steps. These are training-house simulation results, not a claim of generalization. Liquid contact regression: remaining mass .64, .28, 0 after three passes; staying on a spill cannot clean it in one pass. Live visuals use the actual cleaning state and actual neuronal activity.