Reproduce “Hydra: A Real-time Spatial Perception System for 3D Scene Graph Construction and Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nathan Hughes, Yun Chang, Luca Carlone
Hydra incrementally builds 3D scene graphs as robots explore. It limits dense mapping to a moving neighborhood and organizes objects, places, rooms, and buildings into a persistent map. On uHumans2 Office, graph-based loop closure produced approximately twice as many registered matches within 10 cm and 1° error bounds as permissive visual matching; room grouping can miss distinctions in open floor plans.
Key insights
- A "thinking fast and slow" parallelized architecture allows for efficient integration of real-time early/mid-level perception with slower, global optimization tasks.
- Incremental construction of multi-layered scene graphs is achieved using a spatially windowed Euclidean Signed Distance Function (ESDF) and novel graph-based room detection on the place layer.
- Hierarchical, multi-modal loop closure detection leverages visual, object, and place information, while global consistency is maintained by simultaneously optimizing all scene graph layers via an embedded deformation graph.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.