The bot learns a table of expected rewards for actions at each position, heading and altitude. It uses the known simulator to update those values until they settle. Changing obstacles triggers a new solve. This is model-based value iteration, not learning from a camera or trial-and-error reinforcement learning.

Each move costs 1 point, reaching the target earns 40, and a collision ends the run with −30. The planner uses breadth-first search. The untrained table always selects its first action. No safety shield changes the learned bot’s decisions.

Cars use a simplified grid with forward, reverse and quarter-turn moves. Each robot uses a conservative square navigation proxy: 5 m per car tile, 2.5 m per loaded pallet-mover tile, and 1 m for the humanoid, vacuum and drone. Geometry follows this proxy rather than individual limbs, wheels or rotors. Humanoid and pallet robots turn in place; cars turn and move one tile. The pallet mover’s low body carries a load requiring 1.4 m total clearance. Vacuum coverage uses an authored nearest-reachable uncleaned target rule; the learned table navigates to each target.

Obstacle heights are exact synthetic measurements. This experiment does not establish real-world sensing, driving or flight performance.