Abstract

Recent geometric foundation models enable feed-forward inference for SLAM, but their predictions are strongly dependent on the input view set, which leads to geometric inconsistencies and trajectory drift when results are chained over long sequences. Online deployment further exposes a trade-off between the low latency of two-view tracking and the constraint richness of multi-view inference. We introduce UniSim-SLAM, an integrated system that runs lightweight two-view keyframe tracking in the frontend and performs periodic multi-view submap refinement in the backend. To combine predictions defined in heterogeneous local coordinates with inconsistent scales, we formulate a unified multi-level factor graph on Sim(3) that jointly optimizes global keyframe poses and submap poses. The graph integrates temporal view-to-view odometry edges, view-to-submap bridge edges with depth-statistics scale anchoring, and submap-to-submap tie and scale constraints to enforce consistent similarity relations across submaps. Experiments on TUM RGB-D and 7-Scenes show that UniSim-SLAM achieves state-of-the-art accuracy in the uncalibrated setting, reducing trajectory error by 38.5% on TUM RGB-D and 45.9% on 7-Scenes compared to prior best results.

Robots navigating unfamiliar environments must track their position while continuously mapping the space around them. AI has made both tasks faster, but a persistent trade-off remains: using more camera views can improve accuracy, but requires more processing time.

Professor Kyungdon Joo and his research team from the Graduate School of Artificial Intelligence at UNIST have developed UniSim-SLAM, a system that combines rapid camera tracking with periodic multi-view refinement. The approach allows robots to update their position quickly while correcting errors that would otherwise accumulate as they move.

Known as simultaneous localization and mapping (SLAM), the underlying technology enables robots and other autonomous systems to determine their location while reconstructing their surroundings. Recent AI-based approaches can infer 3D geometry directly from camera images, but small differences between successive estimates can compound over long sequences, gradually pushing the reconstructed path away from the robot's actual trajectory.

UniSim-SLAM tackles the problem by giving fast tracking and detailed reconstruction different roles. It compares two camera views at a time to track movement, then periodically brings multiple views together to refine the local 3D map. A unified optimization framework aligns estimates produced at different scales and orientations, keeping the reconstructed environment and camera trajectory consistent over time.

On two standard indoor benchmarks achieved, UniSim-SLAM state-of-the-art results in the uncalibrated setting. With the previous best Compared methods, it reduced trajectory error by 38.5% on TUM RGB-D and 45.9% on 7-Scenes.

The system was also considerably faster than a multi-view alternative evaluated in the study. VGGT-SLAM required 3.41 seconds to produce a result, compared with 0.197 seconds for UniSim-SLAM. The new approach also improved the accuracy of 3D reconstruction.

“UniSim-SLAM brings fast tracking and accurate multi-view analysis into the same system,” said Professor Joo. “This could benefit applications that need reliable localization and real-time 3D reconstruction, from robotics and autonomous driving to augmented reality.”

The study, with Inha Lee of UNIST as first author, was presented at the European Conference on Computer Vision (ECCV 2026), held September 8–12 in Malmö, Sweden. The work was supported by the Ministry of Science and ICT (MSIT) through the Institute of Information & Communications Technology Planning & Evaluation (IITP), including the Artificial Intelligence Graduate School Program at UNIST and programs in collaborative AI and spatial immersive technologies, as well as the InnoCORE program. 

Journal Reference

Inha Lee, Dongjae Jeong, Junhee Lee, and Kyungdon Joo, "UniSim-SLAM: Feed-Forward SLAM with Unified Sim(3) Optimization,"  ECCV'26  (2026).