Open Source Robotics

A curated collection of high-quality open source robotics projects, tools, and software to propel the robotics community forward.

Truncated Signed Distance Function

  • Also called: TSDF, truncated signed distance field
  • Category: Sensing › 3D maps & reconstruction
  • Tags: 3d-reconstruction, mapping, depth-sensing, volumetric, 3d-perception

A voxel map that stores signed distance to the nearest surface only within a thin band around it, fused from many depth images to build dense 3D models.

Definition

A truncated signed distance function (TSDF) represents a 3D surface implicitly on a voxel grid. Each voxel near the surface stores a signed distance to it, positive in front of the surface (in free space as seen from the sensor) and negative behind it, and the surface itself is the zero crossing. Values are clamped, or truncated, to a band of a few voxels on either side. Voxels far from any surface hold no useful distance, which saves computation and avoids corrupting distant regions with noisy measurements.

The volumetric fusion method of Curless and Levoy (1996) introduced the standard update. For each new depth image, every voxel within the camera view is projected into the image, the distance along the camera ray between the voxel and the measured depth is computed and truncated, and the result is merged into a per-voxel weighted running average. Averaging many views cancels sensor noise and fills in surfaces seen from different angles. KinectFusion (Newcombe et al., 2011) showed this running in real time on a GPU with a handheld depth camera, and later systems such as InfiniTAM and Voxblox use voxel hashing or blocks allocated only near surfaces so that large scenes fit in memory. A mesh is extracted with marching cubes, or the zero crossing is found by raycasting for rendering and tracking.

A TSDF is related to, but not the same as, a full signed distance field. Because distances are measured along camera rays and cut off at the truncation band, they are only approximations of true Euclidean distance close to the surface, and they say nothing about space far from it. Planners that need the distance to the nearest obstacle everywhere, such as gradient-based trajectory optimizers, usually compute a Euclidean signed distance field from the TSDF, as Voxblox does. The TSDF is mainly a reconstruction and fusion tool, while the full field is a query structure for planning.

Why it matters

TSDF fusion is the most widely used way to turn a stream of noisy depth images into a clean, dense 3D model. Robots use it to reconstruct objects for grasping, build maps for navigation and produce meshes and point clouds with reliable normals.

Key applications

  • Dense 3D reconstruction from RGB-D and depth cameras
  • Object and scene models for grasp planning
  • Mapping for micro aerial vehicles and mobile robots
  • Building Euclidean distance fields for collision checking
  • Mesh generation for digital twins and simulation assets

In practice

A manipulator with a wrist-mounted depth camera sweeps over a cluttered table along a short arc, capturing around 20 depth frames. Each frame's camera pose comes from forward kinematics and the hand-eye calibration. The frames are fused into a TSDF with 3 to 5 mm voxels and a truncation distance of a few voxels, and the zero crossing is extracted as a point cloud with surface normals. The grasp planner works on this fused cloud rather than on a single frame, which removes holes from occlusion and reduces edge noise. If the arm's calibration is off, the fused surfaces appear doubled or blurred, which is often the first sign that the extrinsic calibration needs repeating.

Common misconceptions

  • A TSDF and a signed distance field are the same thing — a TSDF stores approximate, ray-based distances only near surfaces, while a full field gives distance to the nearest surface everywhere
  • TSDF fusion corrects pose errors — it assumes camera poses are known; bad poses produce doubled or smeared surfaces
  • Smaller voxels always give better models — very small voxels cost memory and receive fewer measurements each, so noise can increase

Who should know this

  • 3D perception and mapping engineers
  • Manipulation engineers building scene models for grasping
  • SLAM and reconstruction researchers
  • Developers of aerial and mobile robot mapping pipelines

Related technologies

  • KinectFusion
  • InfiniTAM
  • Voxblox
  • Open3D
  • Marching Cubes
  • Voxel Hashing

Related terms

  • Signed Distance Field
  • Marching Cubes
  • Depth Camera
  • Point Cloud
  • Surfel
  • Neural Implicit Representation

Related projects

  • LingBot-Map: Streaming 3D reconstruction from one camera: pose and depth for every frame, live
  • Cartographer: Google's real-time 2D and 3D SLAM system for multi-platform robotics
  • OpenDroneMap: Open-source photogrammetry toolkit that turns drone images into maps and 3D models
  • RTAB-Map: Versatile graph-based SLAM supporting RGB-D, stereo, and LiDAR with real-time operation
  • pySLAM: Hybrid Python/C++ visual SLAM pipeline with classic and deep-learning front-ends
  • LeGO-LOAM: Classic lightweight, ground-optimized lidar odometry and mapping baseline (IROS 2018)

This website requires JavaScript to function properly. Please enable JavaScript to explore our collection of open source robotics projects.