TL;DR: We present Elevator-VIGS, a visual-inertial 3D Gaussian Splatting SLAM system that keeps tracking and mapping through elevator rides, achieving state-of-the-art tracking and rendering performance on real-world and simulated elevator sequences.

We present Elevator-VIGS, a visual-inertial 3D Gaussian Splatting SLAM system that keeps tracking and mapping through elevator rides. Inside a moving elevator, the two sensors are in conflict. The camera sees only the robot's motion relative to the elevator, while the IMU senses that motion plus the elevator's motion relative to the world. This conflict is challenging for existing visual-inertial estimators. If vision dominates, the estimator tracks only the robot's motion within the elevator and misses the elevator's rise, and if the conflict remains, the estimator diverges. We observe that the conflict comes from forcing both observations into a single coordinate frame. We instead estimate the robot's pose in the elevator's coordinate frame, and the elevator's motion relative to the world as a per-keyframe transport state, the elevator's rise and vertical velocity, within dense visual-inertial bundle adjustment. Elevator-VIGS detects rides zero-shot with a vision-language model and a depth network, and constrains the transport state at the departure and the arrival. We record real-world and simulated elevator sequences. On these sequences, Elevator-VIGS achieves state-of-the-art tracking and rendering performance. On four elevator-free public benchmarks it keeps the state-of-the-art performance of VIGS-SLAM.

System overview. Elevator-VIGS takes RGB frames and IMU readings as input, tracks the camera online and trains a Gaussian map in parallel, extending VIGS-SLAM with the components outlined in red. Each keyframe, selected by optical flow and initialized through IMU preintegration, holds VIGS-SLAM's 15-dimensional state of pose, velocity and bias together with its disparities. In the frontend, a ConvGRU refines the dense visual correspondences iteratively, and visual-inertial bundle adjustment optimizes the sliding window at every new keyframe. The ride detector ❺ arms on a semantic and a geometric cue from the RGB frames, and the IMU detects the departure and the arrival within the armed window. From the departure until the fold, each keyframe's position and velocity decompose into an elevator-frame part and the transport state ❶, the elevator's rise and vertical velocity. The solve then widens to 17 dimensions, the departure and arrival constraints ❷ constrain the transport state, and the update projection ❸ preserves the transport state through every optimization. Once the keyframes from the elevator leave the sliding window, the deferred fold ❹ writes the rise into the poses and returns them to the world frame. The mapper initializes Gaussians from each new keyframe's depth and refines the map with color and depth rendering losses. Loop closure proposes keyframe pairs from optical flow, a same-floor gate ❻ rejects pairs on different floors, and each detected loop triggers pose-graph bundle adjustment and a Gaussian update.

We release the Elevator-VIGS dataset, 16 real-world elevator sequences with a laser-measured floor-to-floor height as reference and 16 simulated elevator sequences with ground truth through the ride. The data is available for download here.