Mobile-reader trajectories strongly affect phase-based Radio Frequency Identification (RFID) localization because measurement geometry, noise, multipath, and packet loss determine the information available to the estimator. This paper presents a tabular Q-learning method whose reward is computed from the RFID localization error generated along each candidate antenna motion rather than from geometric coverage alone. The method is evaluated in two simulated 3×3 m shelf regions using a 10 cm grid and one hundred Monte Carlo trials. Under matched sensing budgets of 5, 25, and 50 antenna samples, the proposed method reduces the mean error by approximately 72–99% relative to two non-learning baselines, with lower medians and dispersions. Q-table variation decreases substantially after the initial training stage, although later fluctuations remain. The contribution is an interpretable RFID-aware local trajectory-selection layer supported by matched-budget baseline comparisons. The validation is restricted to simulated local regions; hardware transfer and full-warehouse deployment remain subjects for subsequent investigation.