UniMapping: Unified SLAM Framework for Map-Centric Embodied Perception

Simultaneous Localization and Mapping (SLAM) is increasingly expected to provide reusable spatial representations for downstream perception. However, existing approaches often struggle with scale-consistency and producing maps that lack the geometric fidelity required for reliable perception. We propose UniMapping, a unified SLAM framework that constructs a persistent neural-descriptor map from multimodal observations. We introduce a Spatial-Aware Deformable Transformer that injects explicit geometric inductive bias to ensure scale-invariant feature extraction, alongside a Spatial Fusion strategy that decouples feature aggregation from temporal sequences. Extensive experiments on both indoor and outdoor benchmarks demonstrate competitive SLAM performance. Notably, our method significantly enhances downstream tasks (mAP +3.1% and mIoU +7.1%) by leveraging accumulated multi-view context.

Robots and autonomous systems need to understand both where they are and what is around them. Today, these two abilities are often handled by separate systems: one builds a map for navigation, while another recognizes objects and scenes. Because they do not effectively share information, robots may repeatedly process the same observations and struggle in crowded or partially blocked environments. In this work, we introduce UniMapping, a system that allows robots to build a shared memory of their surroundings while also using that memory to better understand the environment. Instead of only analyzing the current camera view, UniMapping combines information collected over time into a consistent 3D map. This helps the robot remember previously seen areas and use past observations to improve scene understanding. We tested UniMapping in both indoor and outdoor environments and found that it improves not only localization accuracy, but also tasks such as object detection and scene segmentation. Our results suggest that shared spatial memory can provide a stronger foundation for future embodied AI systems, including autonomous robots and self-driving vehicles.