ETH Zurich and Google's DiskChunGS Maps Kilometer-Scale Scenes via Chunked Disk Streaming
rsasaki0109 · x · 2026-10-11
Researchers at ETH Zurich and Google Zürich present DiskChunGS (RA-L 2026), tackling the fundamental scalability bottleneck of 3DGS SLAM: GPU memory caps reconstruction at small scenes.
Approach
- Out-of-core chunk-based architecture: scenes are partitioned into spatial chunks; only active regions stay in VRAM while inactive areas stream between disk and GPU
- Integrates with existing SLAM frameworks for pose estimation and loop closure, enabling globally consistent reconstruction at scale
Results
- Uniquely completes all 11 KITTI sequences without memory failures while achieving superior visual quality
- Validated on indoor scenes (Replica, TUM-RGBD), urban driving (KITTI), and resource-constrained Nvidia Jetson platforms, with ROS integration
The authors argue algorithmic innovation can overcome the memory constraints that limited previous 3DGS SLAM methods; paper, video and GitHub are available.
More from Embodied
- Google DeepMind Robotics Hiring Research Scientist in Pretraining & Data Quality in London — bousmalis · 2026-10-11
- GPT-6 Astra for Robotics: a curated collection of papers, reports and evaluations — NielsRogge · 2026-10-11
- Optimus V3 faces twin hurdles: screw-driving precision and parts supply crunch — JOBhakdi · 2026-10-11
- Lab automation only handles cookie-cutter assays — the long tail is what stifles innovation — anshulkundaje · 2026-10-11
- Stanford's Danfei Xu: Robotics Bottleneck Is the Controller and Data, Not Hardware or Models — ZeYanjie · 2026-10-11
- 2027 Predictions: $3,000 Unitree Humanoid, AI-Discovered Theorem in a Journal, ARC-AGI-3 at 50% — flowersslop · 2026-10-11