SLAMFormer-∞ Infinite SLAM Transformer for Unbounded Frontend and Backend Processing

Zhijian Fang^* Weicheng Zheng^* Yijun Yuan^*† Weibang Wang Zhuoguang Chen Chang Sun Junhao Huang Kenan Li Minghui Qin Hang Zhao^†

IIIS, Tsinghua University ^* Equal contribution ^† Corresponding author

Watch demo Paper Explore method BibTeX

Unbounded sequence RGB only

Self-collected 17 kilometer city-scale drive with a satellite route and consistent SLAMFormer-Infinity reconstruction. — Self-collected city-scale drive 17 km · 45 min

Memory conditions keep active predictions in a local coordinate system.

PGGO jointly refines long-range camera poses and dense scene geometry.

See it in motion

Demo

SLAMFormer-∞ processes streaming monocular RGB, maintains an online dense map, and revisits distant regions through a globally connected pose-geometry graph.

SLAMFORMER_INFINITY_DEMO.MP4 RGB → POSE + DENSE GEOMETRY

Watch online SLAM unfold, sequence by sequence.

Each clip shows the online SLAM system building poses and dense geometry on a different KITTI drive.

KITTI_00.MP4 ONLINE SLAM · RGB → POSE + MAP

KITTI Sequence 00 Streaming online reconstruction

ONLINE RUN

AUTO ADVANCE · 20 SEC SEQUENCE 00 / ONLINE RUN

Long-range SLAM

Existing geometric transformers face a structural choice: learn a globally consistent pose and map inside a bounded range, or extend to long sequences by aligning poses and stitching mostly fixed local geometry. SLAMFormer-∞ removes that trade-off.

Strong joint reasoning, bounded by training trajectories.

First-frame-anchored global prediction ties coordinate behavior and attention state to the sequence lengths seen during training.

Scalable pose correction, but geometry is largely stitched.

Long-range alignment can recover a coarse trajectory while local pointmaps remain fragmented or misaligned.

SLAMFormer-∞

Bounded local computation. Unbounded global refinement.

Memory-conditioned inference preserves efficient local processing while PGGO jointly optimizes pose and dense geometry over long-range graph connections.

Comparison of bounded SLAM-Former, pose-aligned VGGT-Long, and memory-conditioned SLAMFormer-Infinity. — A memory condition defines the coordinate system and scale for each target chunk, allowing the same transformer to reason locally and refine globally.

One transformer, three operating rhythms.

Frontend, local backend, and global backend share the same geometric prior. Their difference is the conditioning neighborhood and when refinement is triggered.

Streaming

Conditional frontend

Detect keyframes and incrementally estimate poses and pointmaps from bounded local context.

Periodic

Local backend

Refine the latest fixed-size window and update the memory used by subsequent frontend inference.

Loop / sequence end

Global PGGO

Iteratively update camera poses and dense pointmaps together across the global graph.

City-scale qualitative results

The central result is geometric: global alignment should not stop at the camera path. SLAMFormer-∞ updates the pointmaps with the poses, yielding more coherent local structures along kilometer-scale routes.

City-scale reconstruction comparison on KITTI sequences and a self-collected 17 kilometer drive. — KITTI 05 and 09 plus a self-collected 17 km urban drive. VGGT-Long performs global pose alignment; SLAMFormer-∞ jointly refines pose and geometry and maintains a consistent map where the baseline collapses.

Sequence explorer

Select a KITTI sequence, then orbit, zoom, and pan each point cloud independently: VGGT-Long on the left and SLAMFormer-∞ on the right.

VGGT-Long SLAMFormer-∞ drag: orbit · wheel: zoom · right-drag: pan

VGGT-LongGlobal pose alignment

Loading point cloud…

SLAMFormer-∞Joint pose–geometry refinement

Loading point cloud…

Route Length5067 m

Sequence 02 · KITTI Dataset

Monocular RGB Calibration-free Dense pointmaps

Competitive tracking, stronger long-range geometry.

We foreground the behavior the method was designed for. The numbers below summarize the corresponding tracking gains without reproducing full benchmark tables.

KITTI full sequences

26.358 m 23.011 m

Average ATE RMSE · VGGT-Long → SLAMFormer-∞

Waymo urban segments

1.996 m 1.813 m

Average ATE RMSE · VGGT-Long → SLAMFormer-∞

7-Scenes

0.068 m 0.046 m

ATE RMSE · VGGT-SLAM → SLAMFormer-∞

Cleaner geometry beyond the metrics.

On Replica, the fine stage reduces local drift and cleans fragmented surfaces. This qualitative difference is more informative than the modest aggregate score change.

Replica reconstruction before and after the PGGO fine stage. — Reference RGB · before fine · after fine

Citation

@misc{slamformerinfinity,
  title={SLAMFormer-$\infty$: Infinite SLAM Transformer for Unbounded Frontend and Backend Processing},
  author={Zhijian Fang and Weicheng Zheng and Yijun Yuan and Weibang Wang and Zhuoguang Chen and Chang Sun and Junhao Huang and Kenan Li and Minghui Qin and Hang Zhao},
  year={2026},
  eprint={2608.03429}
}