Splat Map

A splat map is typically an RGBA texture, where each color channel (Red, Green, Blue, Alpha) independently controls the intensity of a specific surface material (e.g., grass, rock, dirt, sand) at a given texel. The shader samples these channels and blends between the corresponding material textures based on their relative strengths. This provides fine-grained, per-pixel control over material distribution, allowing for complex transitions like rocky outcrops in a grassy field or muddy paths through sand.

In simulation for robotics and machine learning, splat maps are often generated procedurally to create vast, varied training terrains. This is a key component of Domain Randomization. Algorithms can:

• Vary material distributions to create diverse friction and visual properties.
• Use noise functions (like Perlin or Worley noise) to create natural-looking patterns.
• Randomize color values within channels to produce endless unique terrain instances. This variability forces trained models (e.g., robotic locomotion policies) to become robust to a wide range of ground conditions, facilitating better sim-to-real transfer.

Beyond visual appearance, splat map channels can be linked to physical parameters in a physics simulation engine. For example:

• The 'rock' channel intensity might correlate with a higher friction coefficient.
• The 'mud' channel could reduce traction and increase damping.
• The 'grass' channel might trigger specific sound effects or particle systems. This allows a single texture to define both the visual and physical characteristics of a terrain, ensuring consistency between what a robot 'sees' and how its actuators interact with the ground, which is critical for high-fidelity simulation.

Splat maps are processed by a specialized terrain shader in real-time graphics pipelines. This shader performs the core operations:

1. Sampling: Reads the RGBA values from the splat map for the current fragment.
2. Texture Fetching: Samples the corresponding albedo, normal, and roughness maps for each material.
3. Blending: Uses the channel weights to lerp (linearly interpolate) or use more advanced blending functions between the material properties.
4. Output: Combines the results into the final surface color, normal, and other PBR (Physically Based Rendering) material outputs. This is highly efficient on modern GPUs.

In terrain authoring tools (e.g., Unity Terrain, Unreal Landscape, World Machine), the splat map is the primary data layer for painterly terrain editing. Artists use brushes to 'paint' materials directly onto the landscape, with the tool writing the brush intensity to the appropriate texture channel. These maps are stored as assets and can be exported for use in other engines or for procedural post-processing. They are often stored at a lower resolution than the final diffuse textures to save memory.

Splat maps work in concert with other core terrain data structures to define a complete environment:

• Heightmap: Defines the terrain's geometry (elevation). Splat materials are projected onto this 3D surface.
• Normal Map: Can be derived from the heightmap or painted separately to add small-scale detail.
• Global Masks: A separate splat map might be used to control biome-level transitions (e.g., forest vs. desert), which then uses local splat maps for material detail within that biome. This layered approach allows for complex, multi-scale environment generation essential for creating realistic and diverse simulation training grounds.

The core function of a splat map is to control the distribution and blending of multiple terrain materials (e.g., grass, dirt, rock, sand) across a landscape. Each color channel (Red, Green, Blue, Alpha) typically represents the intensity of a specific material. The shader samples these channels to lerp (linearly interpolate) between material textures, creating seamless transitions and complex, realistic ground cover essential for high-fidelity simulation environments.

In robotics and autonomous vehicle training, splat maps serve as a ground-truth source for semantic segmentation. By encoding material types into distinct channels, they automatically generate pixel-perfect labels identifying traversable terrain (grass, pavement) versus obstacles (rock, water). This is critical for:

• Training perception models in simulation without manual labeling.
• Evaluating and transferring navigation policies by providing a clear, programmatic definition of the environment's physical properties.

Beyond visuals, splat map channels can drive physical properties across a terrain surface. The intensity value for a material can modulate parameters such as:

• Friction coefficients for wheeled or legged robots.
• Audio footstep sounds (e.g., grass vs. gravel).
• Particle emission (dust on dirt, splashes in mud). This creates a unified data layer where visual appearance directly correlates with simulated physical behavior, enhancing the consistency of the training environment for embodied AI.

Splat maps are often generated algorithmically as part of procedural terrain systems. Using rules, noise functions (like Perlin or Worley noise), and erosion simulations, engines can automatically paint materials based on:

• Slope angle: Rock on steep inclines, soil on gentle slopes.
• Altitude: Snow at high elevations, sand at low.
• Moisture maps: Lush grass near water, dry dirt farther away. This enables the creation of vast, varied, and logically consistent training terrains for Domain Randomization without manual artist intervention.

Splat maps enable efficient rendering of complex terrain. Instead of using a single massive texture or thousands of individual material tiles, the splat map allows the GPU to blend between a small set of tiling material textures sampled via triplanar mapping. This approach:

• Drastically reduces texture memory usage and bandwidth.
• Eliminates visible tiling repetition.
• Maintains high visual detail at any scale, which is essential for real-time simulation where rendering performance directly impacts training speed.

Modern splat maps are designed for Physically Based Rendering (PBR) pipelines. Each material channel typically references a set of PBR texture maps (Albedo, Normal, Roughness, etc.). The splat map controls the blending of these full material sets, ensuring that lighting interactions—like specular highlights on wet rock or diffuse scattering on dry grass—are blended correctly. This integration is key for achieving the photorealism required in high-end digital twins and simulation used for computer vision training.

A heightmap is a grayscale image or 2D array where the intensity of each pixel represents the elevation at that point. It is the primary data structure for defining terrain geometry before surface materials are applied with a splat map.

• Core Function: Encodes the 3D shape (mesh) of a landscape.
• Data Format: Typically a single-channel (grayscale) image or a 2D array of float values.
• Workflow Relationship: The heightmap generates the terrain mesh; the splat map then 'paints' materials onto that mesh based on slope, height, or other rules.

Procedural Terrain Generation is the algorithmic creation of landscape geometry and features using mathematical functions. Splat maps are automatically generated as part of this process to assign materials.

• Key Algorithms: Uses Perlin Noise, Voronoi Tessellation, and hydraulic erosion simulations.
• Role of Splat Maps: Algorithms analyze the generated heightmap (e.g., identifying steep slopes as 'rock', flat areas as 'grass') to produce a corresponding splat map.
• Benefit: Enables creation of vast, unique, and realistic training environments for sim-to-real transfer without manual artist work.

Physically Based Rendering (PBR) is a shading model that uses real-world material properties. Splat maps feed into the PBR pipeline to control which PBR material (with its albedo, roughness, normal maps) is rendered on each part of the terrain.

• Material Properties: A PBR material for 'mud' will have specific albedo (color), roughness (wetness), and normal (bumpiness) maps.
• Splat Map Integration: The splat map's red channel might blend the 'mud' material, while the green channel blends a 'grass' material, with the PBR shader combining their properties realistically.

Domain Randomization is a training technique that varies simulation parameters to improve model robustness. Splat maps are a key parameter for visual randomization.

• Application: Between training episodes, the system can randomize the colors in the splat map, swap the material textures it references, or change the distribution rules (e.g., make 'snow' appear at lower altitudes).
• Goal: Prevents the AI policy from overfitting to specific visual patterns (like a particular grass texture), forcing it to learn the underlying physics of the task, which aids sim-to-real transfer.

A Shader Graph is a visual tool for authoring shaders. It is commonly used to create the complex terrain shader that interprets a splat map and blends multiple materials in real-time.

• Visual Workflow: Nodes for texture sampling (splat map, material maps), math operations for blending, and PBR lighting calculations are wired together visually.
• Function: The graph defines how the RGBA values of the splat map control the weighted blend of four (or more) texture sets (diffuse, normal, roughness) on the terrain surface.

Texture Synthesis is the algorithmic generation of large, seamless texture images from a small sample. This is used to create the high-resolution, tileable material textures (e.g., for rock, sand) that are referenced and blended by a splat map.

• Problem Solved: Manually created textures can be small and repetitive. Synthesis creates vast, non-repeating variations.
• Pipeline Integration: A procedural terrain system might use Wave Function Collapse or neural networks to synthesize unique material variations for different regions defined by the splat map, increasing environmental diversity for AI training.