INTEGRATING ARTIFICIAL INTELLIGENCE AND SLAM FOR AUTONOMOUS MOBILE ROBOTS: A SEMANTIC, DYNAMICS-AWARE ARCHITECTURE AND A CONTROLLED LOCALIZATION STUDY

Simultaneous localization and mapping (SLAM) estimates the pose of an autonomous mobile robot where satellite positioning is unavailable, but classical estimators assume a static world. This paper proposes a layered architecture that couples SLAM, semantic perception and navigation through bidirectional data flows and an independent safety layer, and formalises it, including extended Kalman filter (EKF) SLAM equations with analytic Jacobians, semantic masking as a measurement-selection rule and evaluation metrics. A Monte Carlo simulation (40 seeds per condition) quantifies how much masking improves localization when 0 to 40 % of 40 landmarks in a 30 m by 30 m world move at 0.2 to 1.0 m/s; the detector is a confusion model with recall r and false-positive rate f. In this simulation the absolute trajectory error (ATE) of a static-world EKF-SLAM rose from 8.3 cm to 3.7 to 6.4 m with 10 to 40 % moving landmarks. Masking alone at recall 0.9 left 1.3 to 2.6 m. Masking with chi-square gating gave 8.7 to 10.8 cm (oracle detector 8.6 to 10.5 cm) and gating alone 9.0 to 11.9 cm; the combination was better than gating alone by 3 to 10 % on average, and for creeping landmarks (0.02 to 0.1 m/s) its mean ATE was 9.9 cm against 22.5 cm, a difference concentrated in a minority of seeds. Masking pays off only with near-perfect recall, or when combined with gating or temporal voting. Limitations are the abstract detector and independent detector errors; hardware validation remains future work.