<p>Bed-chair robots, as a specialized class of intelligent wheelchairs, face persistent challenges in achieving robust localization within complex indoor environments. These challenges stem from sparse visual textures in featureless corridors, dynamic obstacles in crowded areas, abrupt motion changes, and varying lighting conditions. In this work, we propose multi-sensor fusion (MS_Fusion), an adaptive multi-sensor fusion SLAM system that leverages the structured regularities of the Manhattan World assumption within a factor graph optimization framework. By dynamically adjusting the weights of sensor inputs—including visual features, inertial measurement units (IMUs), wheel odometry, and geometric constraints—our system can adapt to real-time environmental and motion conditions. This adaptive weighting mechanism ensures high localization accuracy and robustness even in degenerate scenarios, such as featureless walls or odometric slippage. To further enhance real-time performance, we introduce an adaptive sliding window strategy and an active map point selection policy, which together reduce computational overhead without compromising accuracy. Extensive experiments on public datasets and proprietary benchmarks demonstrate that MS_Fusion outperforms state-of-the-art low-cost SLAM systems, achieving superior robustness, precision, and efficiency in diverse and challenging indoor scenarios. Beyond technical contributions, this work has practical implications for improving the autonomy and reliability of assistive robots in real-world applications. The codebase and datasets will be made publicly available incrementally in the future to facilitate further research and development on GitHub.</p> Graphical abstract <p></p>