Voxblox: Building 3D Signed Distance Fields for Planning

Truncated Signed Distance Fields (TSDFs) have become a popular tool in 3D reconstruction, as they allow building very high-resolution models of the environment in realtime on GPU. However, they have rarely been used for planning on robotic platforms, mostly due to high computational and memory requirements. We propose to reduce these requirements by using large voxel sizes, and extend the standard TSDF representation to be faster and better model the environment at these scales. We also propose a method to build Euclidean Signed Distance Fields (ESDFs), which are a common representation for planning, incrementally out of our TSDF representation. ESDFs provide Euclidean distance to the nearest obstacle at any point in the map, and also provide collision gradient information for use with optimization-based planners. We validate the reconstruction accuracy and real-time performance of our combined system on both new and standard datasets from stereo and RGB-D imagery. The complete system will be made available as an open-source library called voxblox.