Abstract:

Fine-grained mapping and intelligent inspection of underground spaces are fundamental to autonomous robotic operation, emergency rescue, and safety maintenance. However, existing 3D Gaussian Splatting SLAM (3DGS-SLAM) methods mainly emphasize geometric reconstruction and photorealistic rendering, while their semantic understanding capability remains limited in complex underground environments. Moreover, the storage and optimization of high-dimensional semantic features introduce substantial memory overhead, and degraded observations caused by low illumination and incomplete depth measurements often lead to inconsistent semantic representations across multiple views. To address these issues, this paper proposes an open-vocabulary semantic-enhanced 3DGS-SLAM method for complex underground spaces. First, a multi-scale open-vocabulary semantic feature field is constructed using SAM (segment anything model) and CLIP (contrastive language-image pretraining), together with depth completion and confidence estimation. An autoencoder is further introduced to compress CLIP features, reducing GPU memory consumption while preserving semantic discriminability. Second, a semantic uncertainty quantification model is proposed by integrating an image-enhancement residual prior and a depth-confidence prior. A temporal smoothing mechanism is also introduced to reduce the semantic weights of low-quality observations, thereby mitigating multi-view semantic inconsistency. Finally, photometric, depth, Gaussian-scale regularization, illumination-stability, and uncertainty-weighted semantic losses are combined to form a joint optimization objective for geometry, appearance, and semantics, while supporting open-vocabulary semantic queries. To validate the effectiveness of the proposed method, experiments were conducted in multiple representative underground spaces using a self-integrated mobile robotic platform. The proposed method achieves a peak signal-to-noise ratio (PSNR) of 23.0 dB and a structural similarity index measure (SSIM) of 0.88 in underground spaces, outperforming representative methods such as SplaTAM and GS3LAM. On the ScanNet dataset, it achieves an open-vocabulary segmentation mIoU of 39.08%, showing clear improvement over competing methods. The results demonstrate that the proposed method can achieve efficient and stable open-vocabulary semantic mapping in complex underground environments, providing fine-grained semantic support for intelligent underground inspection and human-robot interaction.

Key words: underground spaces, 3DGS-SLAM, open-vocabulary semantic mapping, semantic uncertainty modeling, mobile robots