DiskChunGS: Large-Scale 3D Gaussian SLAM Through Efficient Disk Chunking

Casimir Feldmann1, Maximum Wilder-Smith1, Vaishakh Patil1, Micheal Oechsle2, Micheal Niemeyer2, Keisuke Tateno2, Marco Hutter1
1ETH Zürich 2Google Zürich
Accepted to RA-L 2026
Paper arXiv Video Coming Soon

Map kilometer-scale environments with photorealistic 3D Gaussians by streaming scene chunks between disk and GPU memory.

Abstract

Recent advances in 3D Gaussian Splatting (3DGS) have demonstrated impressive results for novel view synthesis with real-time rendering capabilities. However, integrating 3DGS with SLAM systems faces a fundamental scalability limitation: methods are constrained by GPU memory capacity, restricting reconstruction to small-scale environments. We present DiskChunGS, a scalable 3DGS SLAM system that overcomes this bottleneck through an out-of-core approach that partitions scenes into spatial chunks and maintains only active regions in GPU memory while storing inactive areas on disk. Our architecture integrates seamlessly with existing SLAM frameworks for pose estimation and loop closure, enabling globally consistent reconstruction at scale. We validate DiskChunGS on indoor scenes (Replica, TUM-RGBD), urban driving scenarios (KITTI), and resource-constrained Nvidia Jetson platforms. Our method uniquely completes all 11 KITTI sequences without memory failures while achieving superior visual quality, demonstrating that algorithmic innovation can overcome the memory constraints that have limited previous 3DGS SLAM methods.

We present three main contributions to the state-of-the-art:

  1. Out-of-core chunk-based architecture that enables large-scale 3DGS SLAM by partitioning scenes into spatial regions and dynamically managing them between disk and VRAM
  2. Comprehensive evaluation demonstrating state-of-the-art performance across indoor and outdoor datasets without memory failures
  3. Production-ready deployment validated on Jetson and integrated with ROS for robotic platforms

Video

Pipeline

For each keyframe, we estimate depth and place Gaussian primitives based on image content. Frustum culling determines which spatial chunks are visible. These are loaded from disk into VRAM while inactive chunks are evicted. The active Gaussians are then optimized using photometric and depth losses.

Reconstructions

Slide to compare reconstruction quality on KITTI against state-of-the-art methods CaRtGS and On-The-Fly-NVS.

Scene 05

Ground Truth Ours On-The-Fly-NVS CaRtGS

Scene 07

Ground Truth Ours On-The-Fly-NVS CaRtGS

Scene 10

Ground Truth Ours On-The-Fly-NVS CaRtGS

Interactive Chunk Management Visualization

Move your cursor to control the camera frustum (triangle). This demonstrates DiskChunGS's LRU-based chunk caching system.
As you move the frustum, chunks are loaded into a limited VRAM cache. When the cache is full, the least recently used chunks are evicted back to disk.

Visible Chunks - Currently visible and being optimized/rendered
Chunks in VRAM - Recently accessed, ready for quick optimization/rendering
Stored on Disk - Not cached, requires loading from disk if needed

BibTeX

@article{feldmann2025diskchungslargescale3dgaussian,
        title = {DiskChunGS: Large-Scale 3D Gaussian SLAM Through Chunk-Based Memory Management}, 
        author = {Casimir Feldmann and Maximum Wilder-Smith and Vaishakh Patil and Michael Oechsle and Michael Niemeyer and Keisuke Tateno and Marco Hutter},
        journal = {arXiv preprint arXiv:2511.23030},
        year = {2025},
        eprint = {2511.23030},
        archivePrefix = {arXiv},
        primaryClass = {cs.RO},
        url = {https://arxiv.org/abs/2511.23030}
      }, 
}