SAFARI: Semantic–Depth Adaptive SLAM for Visually Challenging Robotic Environments

Abstract Robust simultaneous localization and mapping (SLAM) remains difficult when the visual appearance of a scene changes faster than a fixed SLAM configuration can accommodate. Low texture, reflections, turbidity, partial occlusion, and repetitive structures can simultaneously degrade feature extraction, monocular depth prediction, and map optimization. This paper presents SAFARI, a semantic--depth adaptive SLAM framework that couples dense monocular depth, vision--language scene interpretation, and constrained online parameter adaptation. Depth Anything V2 supplies dense geometric cues, CLIP produces scene-level semantic evidence, and a lightweight language-model controller converts compact perception and optimization statistics into bounded updates of mapping parameters such as keyframe density, optimization iterations, and pruning behavior. A semantic--depth confidence module attenuates unreliable depth in semantically identified difficult regions before the observations are incorporated into the map. The framework is evaluated on the Replica indoor benchmark and on a marine-pool sequence acquired with an underwater remotely operated vehicle (KUMP). On the reported Replica ablation, semantic masking increases reconstruction PSNR from 20.18 to 34.25 dB and reduces depth L1 error from 0.23 to approximately zero at the evaluated frame, while yielding an IoU of 0.8916. On the underwater sequence, the complete pipeline achieves an ATE RMSE of 41.79 cm, RGB PSNR of 19.12 dB, and depth L1 error of 7.79 cm over the evaluated 200-frame sequence. These experiments support the feasibility of context-conditioned SLAM adaptation, while also exposing current limitations in long-horizon validation, underwater semantic segmentation, and computational efficiency.