Elevator-VIGS: Gaussian Splatting SLAM that keeps tracking through elevator rides, zero-shot via VLM
zhenjun_zhao · x · 2026-09-22
A paper by Marc Pollefeys and colleagues tackles visual-inertial Gaussian Splatting SLAM during elevator rides, where the camera sees only robot-relative motion while the IMU senses motion plus the elevator's, causing divergence or missed ascents.
- Key idea: estimate the robot pose in the elevator's coordinate frame and model the elevator's rise and vertical velocity as a per-keyframe transport state inside dense visual-inertial bundle adjustment
- Detects rides zero-shot with a vision-language model plus a depth network, constraining the transport state at departure and arrival
- Achieves SOTA tracking and rendering on real and simulated elevator sequences while preserving VIGS-SLAM SOTA on four elevator-free benchmarks
More from Embodied
- Wayve and Mercedes sign global deal to bring AI Driver to production cars in two years — alexgkendall · 2026-09-22
- MINIX N304-AI mini PC packs Intel Core 3 304 with 24-TOPS NPU — DeliciousBelt9520 · 2026-09-22
- Remi Cadene podcast with JB on the hardest tasks for humanoid robots and dexterity — RemiCadene · 2026-09-22
- Chatbots Were the Warm-Up: Why World Models May Underpin Physical AI — ingliguori · 2026-09-22
- BayesianGS-SLAM: Uncertainty-Aware 3DGS SLAM Uses Bayesian Uncertainty Across Mapping, Tracking and Keyframes — zhenjun_zhao · 2026-09-22
- Object-Path Graphs Unify Open-Vocabulary Semantics and Topological Navigation, Skipping Dense Maps — zhenjun_zhao · 2026-09-22