OREN: Hybrid Octree-Neural Network for Real-Time SDF Mapping
A novel technique known as OREN (Octree Residual Network) integrates octree interpolation with neural network regression to facilitate real-time reconstruction of the Euclidean signed distance function (SDF) from point cloud data. This method overcomes the shortcomings of current volumetric approaches, which are discrete and non-differentiable, as well as neural network techniques that suffer from inefficiency, catastrophic forgetting, memory limitations, and truncated SDF issues. OREN provides non-truncated SDF reconstruction while achieving computational and memory efficiency akin to volumetric methods, all the while ensuring differentiability and precision. This research was made available on arXiv (2510.18999) and is aimed at enhancing robot autonomy in areas such as localization, mapping, motion planning, and control.
Key facts
- OREN stands for Octree Residual Network
- Method combines explicit octree interpolation prior with implicit neural network residual
- Achieves non-truncated Euclidean SDF reconstruction
- Computational and memory efficiency comparable to volumetric methods
- Supports differentiability and accuracy
- Targets robot autonomy: localization, mapping, motion planning, control
- Published on arXiv with ID 2510.18999
- Addresses catastrophic forgetting and memory limitations of neural networks
Entities
Institutions
- arXiv