Signed Distance Field (SDF)

The core property of an SDF is that at any point in space, its value is the shortest Euclidean distance to the surface of a shape. The sign of this value carries critical information:

• Positive values indicate the point is outside the shape.
• Negative values indicate the point is inside the shape.
• A value of zero defines the iso-surface, which is the shape's boundary. This sign allows for trivial inside/outside tests, which are fundamental for collision detection and constructive solid geometry (CSG) operations.

An SDF is an implicit function, Φ(x), that defines a surface as the set of all points where Φ(x) = 0 (the zero-level set). Unlike explicit representations like meshes, the surface is not stored directly but is defined by a rule. This enables:

• Natural handling of topology changes (e.g., merging or splitting shapes) without remeshing.
• Straightforward computation of surface properties like normals, which are given by the normalized gradient (∇Φ/‖∇Φ‖) at the surface.
• Robust Boolean operations (union, intersection, difference) through simple min/max operations on the distance fields.

SDFs can be represented in two primary forms:

• Analytic SDFs are defined by closed-form mathematical formulas for primitive shapes (e.g., spheres, boxes, planes). For a sphere centered at c with radius r, the SDF is simply Φ(x) = ‖x - c‖ - r. These are exact, efficient, and differentiable.
• Discrete SDFs are sampled on a voxel grid or octree, where each grid cell stores a precomputed distance value. This representation is used for complex, arbitrary shapes (like a 3D scan of an object) and enables GPU-friendly processing via 3D textures. The trade-off is memory usage and sampling resolution.

SDFs are central to several advanced computer graphics techniques:

• Ray Marching (Sphere Tracing): An efficient rendering algorithm that uses the SDF's distance value as a safe step size to find surface intersections without expensive geometric tests.
• Anti-aliased Rendering: The distance value provides sub-pixel information, enabling smooth edges and outlines (e.g., for font rendering).
• Procedural Modeling & CSG: Complex models are built by combining simple SDF primitives using smooth Boolean operators, enabling entirely code-driven asset creation.

In robotics and 3D vision, SDFs provide a unified representation for perception and planning:

• Collision Checking: Determining if a robot's configuration is in collision is a fast lookup or evaluation of the SDF sign.
• Motion Planning: The gradient of the SDF can be used to compute repulsive forces for obstacle avoidance in potential field methods.
• 3D Reconstruction: Modern techniques like Neural Signed Distance Functions (Neural SDFs) use multilayer perceptrons to represent SDFs continuously, enabling high-fidelity surface reconstruction from partial point clouds or images.

SDFs are one member of a family of 3D representations, each with trade-offs:

• Vs. Point Clouds: An SDF provides a continuous, complete volumetric field, whereas a point cloud is a sparse set of surface samples.
• Vs. Voxel Grids (Occupancy): A binary occupancy grid only indicates if a cell is full or empty. An SDF provides a continuous distance measure, conveying proximity.
• Vs. Meshes: A mesh is an explicit, lightweight surface representation ideal for rendering but fragile for topology changes. An SDF is implicit and robust but can be more memory-intensive for high-resolution discrete forms.
• Vs. NeRF: Neural Radiance Fields model color and density for view synthesis. A Neural SDF specifically models geometry via distance, often leading to higher-quality surfaces.

SDFs are the core data structure for real-time ray marching, a rendering technique used in shader graphics and demoscene productions. The algorithm samples the SDF along a ray to find surface intersections efficiently.

• Key Advantage: The distance value provides a safe step size, guaranteeing the ray never overshoots the surface, enabling robust and fast rendering of complex, procedurally defined shapes.
• Use Case: Generating complex UI elements, volumetric text, and dynamic visual effects entirely on the GPU without pre-computed meshes.

In robotics simulation and game physics engines, SDFs provide an exceptionally efficient method for proximity and collision queries.

• Proximity Queries: Determining the exact distance between two objects or an object and a point is a constant-time lookup or a fast evaluation.
• Collision Normal & Penetration Depth: The gradient of the SDF at a point gives the surface normal, and the distance value itself provides the penetration depth, which are critical for resolving collisions and applying contact forces.
• Application: Used in model predictive control (MPC) and motion planning to maintain safe distances from obstacles.

SDFs are the standard implicit representation in modern 3D reconstruction pipelines from multiview images or point clouds.

• Neural SDFs: Techniques like Signed Distance Functions (SDFs) in Neural Radiance Fields (NeRF) variations (e.g., NeuS) use a neural network to regress a continuous SDF from images, yielding high-fidelity, watertight meshes via the Marching Cubes algorithm.
• Advantage over Explicit Representations: The implicit SDF naturally handles topological changes and incomplete data, smoothly interpolating missing surfaces.

For mobile robots and manipulators, an SDF of the environment is a powerful tool for global and local planning.

• Distance Gradient for Planning: Algorithms like Gradient Descent can use the SDF's spatial gradient to repel a planned path away from obstacles, creating inherently safe trajectories.
• Integration with Occupancy Grids: An SDF can be efficiently computed from a probabilistic occupancy grid, transforming a binary map into a smooth distance field that planners can use to optimize for clearance.
• Example: The MoveIt motion planning framework can utilize SDFs for collision checking in complex scenes.

The mathematical properties of SDFs enable precise volumetric shape analysis and processing.

• Boolean Operations: Union, intersection, and difference of shapes are trivial and robust using min, max, and negation operations on their respective SDFs.
• Offsetting & Dilating: Creating an expanded (dilated) or contracted (eroded) version of a shape is performed by simply subtracting or adding a constant to the SDF.
• Application: Used in computer-aided design (CAD) for model blending, in medical imaging for tissue thickness measurement, and in additive manufacturing for generating infill structures.

SDFs are famously used for rendering crisp, resolution-independent vector shapes, most notably for font glyphs.

• Mechanism: A 2D SDF texture is generated for each glyph, storing the distance to the outline. During rendering, a fragment shader samples this texture and uses a smoothstep function on the distance to produce perfectly anti-aliased edges at any scale or rotation.
• Performance: This allows an entire font to be stored in a single, small texture atlas and rendered with extreme efficiency, a technique pioneered by Valve and used ubiquitously in game engines and UI frameworks.

A voxel grid is a discrete, grid-based volumetric representation where 3D space is divided into a regular lattice of cubic cells. Each voxel stores data, such as binary occupancy, density, or color.

• Key Difference from SDF: While an SDF stores a continuous distance value, a voxel grid typically stores discrete attributes per cell. SDFs can be stored within a voxel grid structure, where each voxel holds the SDF value at that location.
• Use Case: Efficient for processing with 3D Convolutional Neural Networks (3D CNNs) due to their regular structure. Common in early deep learning for 3D shape classification and segmentation.

A point cloud is a set of discrete data points in a 3D coordinate system, representing the external surfaces of objects as captured by sensors like LiDAR or depth cameras.

• Key Difference from SDF: A point cloud is a sparse, surface-only representation, whereas an SDF is a dense, volumetric field defined throughout space. SDFs are often derived from point clouds through reconstruction algorithms.
• Use Case: The raw output of most 3D sensors. Used for visualization, registration (e.g., ICP), and as input to networks like PointNet.

Neural Radiance Fields (NeRF) represent a 3D scene as a continuous volumetric function, typically a multilayer perceptron (MLP). This function maps a 3D location and viewing direction to a volume density and view-dependent color.

• Key Difference from SDF: NeRF models appearance (color) and density for photorealistic rendering, while a pure SDF models only geometry (distance to surface). Modern extensions like NeuS and VolSDF unify them by using an SDF to guide the density field.
• Use Case: High-fidelity novel view synthesis from a set of 2D images.

An occupancy grid is a probabilistic, discrete spatial representation where each cell (often a voxel) stores the probability that it is occupied by an obstacle.

• Key Difference from SDF: Occupancy is a binary or probabilistic classification (occupied/free), while an SDF provides a metric distance and direction to the nearest surface. An SDF contains strictly more information and can be thresholded to create an occupancy grid.
• Use Case: Fundamental to robotic mapping and navigation (e.g., for path planning in SLAM systems), where knowing free space is critical.

Surface reconstruction is the process of creating a continuous 2D manifold (a triangle mesh) from discrete 3D data like a point cloud or an SDF.

• Key Relationship to SDF: The zero-level set of an SDF (the set of points where distance = 0) implicitly defines the object's surface. Algorithms like Marching Cubes or Dual Contouring can extract a mesh from a discrete SDF volume.
• Use Case: Converting raw sensor data or learned implicit fields into usable 3D models for simulation, CAD, and computer graphics.

A Truncated Signed Distance Field (TSDF) is a practical variant of an SDF where distance values are only stored and considered within a fixed band (truncation distance) around the observed surface.

• Key Advantage: Dramatically reduces memory and computational cost compared to a full SDF, as only relevant near-surface volumes are processed. It robustly fuses multiple depth maps by averaging weighted updates.
• Use Case: The core representation in real-time 3D reconstruction and dense SLAM systems like KinectFusion, enabling live fusion of depth camera data into a consistent volumetric model.