S2GO

Overview

S2GO (Streaming Sparse Gaussian Occupancy) is a new approach to 3D scene understanding that builds a lightweight, streaming 3D occupancy map using a small set of learned 3D queries that evolve over time. It maintains a compact query-based world state, enabling dense, high-quality occupancy prediction up to 5.9× faster than prior methods and supporting long-horizon, camera-only, real-time perception. This blog post explains the motivation, method, and integration of S2GO into modern autonomy stacks.

Key features

• ▸Streaming sparse Gaussian occupancy for real-time 3D perception
• ▸Up to 5.9× faster than prior methods
• ▸Camera-only, single GPU real-time operation
• ▸Compact query-based world state (approx. 1000 queries)
• ▸Geometry-first pretraining with LiDAR denoising and rendering
• ▸Temporal consistency across frames with persistent queries
• ▸State-of-the-art on nuScenes and KITTI-360 benchmarks