Global Map

The global map is a single, cohesive model of the environment, integrating all observed data into one consistent frame of reference. It is built by fusing multiple local submaps—temporary maps created from sequential sensor observations—into a unified whole. This process requires solving the data association problem to correctly match features and landmarks across different submaps and time steps.

A primary function of the global map is to maintain global consistency, correcting errors that accumulate over time. This is achieved through loop closure detection, where the system recognizes a previously visited location. When a loop is closed, a pose graph optimization (like g2o or GTSAM) is triggered, adjusting all past robot poses and map landmarks to minimize total error, eliminating drift and ensuring the map is metrically accurate over large scales.

Unlike transient local maps, the global map is persistent across sessions and must be scalable to represent environments from a single room to an entire city. This requires efficient data structures and multi-resolution representations.

• Hierarchical Mapping: Uses coarse-to-fine representations (e.g., an occupancy grid at city scale, detailed meshes for local interaction).
• Incremental Updates: Allows for continuous expansion without full recomputation.
• Long-Term Data Management: Involves strategies for pruning redundant data or summarizing old observations to manage memory.

Modern global maps extend beyond pure geometry to include semantic labels and object-level understanding. This is achieved by integrating outputs from semantic segmentation and 3D object detection.

• Semantic Layers: Map elements are tagged with classes (e.g., wall, door, chair, road).
• Instance-Aware: Tracks individual objects (e.g., car_17, desk_2) over time.
• Relational Information: Can encode relationships (e.g., chair is on floor, near desk).

This enrichment enables higher-level reasoning for tasks like "navigate to the kitchen table" or "avoid all dynamic obstacles."

A robust global map is built by fusing data from multiple complementary sensors to overcome the limitations of any single modality.

• Visual (Cameras): Provide rich texture and feature points for Visual SLAM.
• Depth (LiDAR/ToF/RGB-D): Supply precise 3D point clouds for geometric accuracy.
• Inertial (IMU): Offers high-frequency motion data for stability during rapid movement or visual degradation, as used in Visual-Inertial Odometry (VIO).
• GPS (Outdoor): Provides absolute global positioning to anchor the map and mitigate long-term drift.

Fusion typically occurs within a factor graph or Kalman filter framework.

The internal structure of a global map can take several forms, each with trade-offs in precision, memory, and query speed.

• Sparse Feature Maps: Store only distinct landmarks (e.g., ORB, SIFT features). Efficient for localization but lacks dense geometry. Used in systems like ORB-SLAM.
• Dense Volumetric Maps: Represent space as a 3D grid of voxels, often storing an occupancy probability or Truncated Signed Distance Function (TSDF). Provides complete surface geometry but is memory-intensive.
• Hybrid Representations: Combine sparse pose graphs with local dense submaps (e.g., Kimera, ElasticFusion).
• Neural Implicit Maps: Use a coordinate-based neural network (like a variant of NeRF or an implicit surface representation) to encode geometry and appearance continuously. Highly compact but computationally expensive to query.

A global map enables persistent world-locked content in mixed reality. Virtual objects placed by a user remain anchored in the correct physical location across multiple application sessions. This requires the map to be stored, recalled, and updated. Key technologies include:

• Spatial Anchors for precise virtual object placement.
• World Meshes for environmental occlusion and physics.
• Frameworks like ARKit and ARCore that manage persistent cloud maps.

For mobile robots and drones, the global map is the core navigation graph. It integrates semantic segmentation data (identifying floors, doors, obstacles) to enable long-term path planning and obstacle avoidance. The system continuously localizes the robot within this map using Visual SLAM or LiDAR SLAM. This allows for:

• Multi-room and multi-floor autonomous operation.
• Efficient re-planning when encountering dynamic obstacles.
• Loop closure to correct cumulative odometry drift over large areas.

Global maps form the geometric and semantic backbone of digital twins for factories, cities, or buildings. They are constructed by fusing data from drones, mobile scanners, and fixed sensors. The resulting map is not just a 3D model but a queryable database containing:

• Asset locations (machinery, utilities).
• Structural semantics (walls, windows, conduits).
• Historical change tracking for maintenance and simulation. This enables predictive analytics and virtual walkthroughs.

In warehouses, ports, or construction sites, a shared global map is essential for coordinating a heterogeneous fleet of robots and autonomous guided vehicles (AGVs). This map acts as a single source of truth for:

• Dynamic task allocation and zone management.
• Collision avoidance and traffic flow optimization.
• Real-time updates on environmental changes (e.g., a spilled pallet). Agents localize themselves within this shared map, enabling decentralized yet coherent behavior.

Global maps generated from Visual-Inertial Odometry (VIO) on drones or handheld scanners create accurate as-built models of infrastructure like bridges, pipelines, or cell towers. The map's global consistency allows for:

• Precise measurement of deformations or corrosion over time.
• Change detection between successive surveys.
• Annotation of defects directly onto the 3D model for repair crews. The process often uses Bundle Adjustment on collected imagery to achieve survey-grade accuracy.

Beyond GPS, a pre-built global map enables camera-based localization indoors and in urban canyons. A device captures an image, and its features are matched against a cloud-based global map to determine its 6DoF pose with centimeter-level accuracy. Applications include:

• Indoor navigation in airports and malls.
• Asset tracking in large industrial facilities.
• Augmented reality guidance for field technicians, overlaying instructions directly on equipment.

A local map is a temporary, high-fidelity representation of the immediate surroundings, built from recent sensor data. It is the fundamental building block fused into the global map.

• Function: Provides dense, accurate geometry for real-time navigation and obstacle avoidance.
• Characteristics: Often exists as a point cloud or voxel grid and is continuously updated.
• Relationship to Global Map: Multiple local maps are aligned and merged using techniques like Iterative Closest Point (ICP) to form the consistent global representation.

A pose graph is a sparse optimization framework that represents the global map's topological structure. Nodes are estimated camera or robot poses, and edges are spatial constraints between them.

• Core Mechanism: Enables efficient loop closure; when a revisited location is recognized, a new constraint (edge) is added, allowing the system to correct drift across the entire graph.
• Efficiency: Unlike adjusting every point in a dense map, optimizing the pose graph corrects the underlying skeleton of poses, making large-scale mapping computationally feasible.
• Output: The optimized graph provides the corrected 6DoF pose for each node, which is used to transform and align all local map data into a globally consistent frame.

Loop closure is the critical process where a system recognizes it has returned to a previously mapped area, enabling it to correct accumulated odometric error in the global map.

• Problem it Solves: Dead reckoning from sensors like Visual-Inertial Odometry (VIO) inevitably drifts over time.
• Process: Involves place recognition (matching current visual features to a database) followed by pose graph optimization to "pull" the map and trajectory into alignment.
• Result: Without loop closure, a global map would be a collection of misaligned submaps; with it, the map becomes a single, consistent coordinate frame.

A semantic map augments the geometric global map with object-level labels and meaning. It answers "what" is in the environment, not just "where" surfaces are.

• Construction: Built by projecting the results of semantic segmentation from 2D images onto the 3D map geometry.
• Applications: Enables higher-level reasoning for robots ("navigate to the kitchen") and AR applications ("attach the virtual screen to the wall").
• Data Structure: Often stored as a labeled voxel grid or a set of object instances with bounding volumes and class IDs within the global coordinate system.

A spatial anchor is a persistent point of interest attached to a specific location in the global map. It allows virtual content to be precisely recalled across multiple application sessions.

• Persistence: The anchor's pose (position and orientation) is stored relative to the global map's coordinate system.
• Cloud Relocalization: Systems like ARKit and ARCore can save anchors to a cloud service. When a user returns, the device downloads the anchor and uses visual features to relocalize itself within the stored global map to find the anchor's exact world position.
• Use Case: Essential for multi-user AR experiences and persistent digital overlays in enterprise and industrial settings.

An occupancy grid is a probabilistic, discretized representation of space used for navigation-centric global maps. Each cell (or voxel) stores the probability that it is occupied by an obstacle.

• Contrast with Dense Maps: Unlike a detailed point cloud or mesh, an occupancy grid sacrifices fine geometry for efficient path planning and collision checking.
• Sensor Fusion: Updated by fusing successive depth measurements from LiDAR, stereo cameras, or depth sensors using Bayesian filtering.
• Application: The foundational map representation in robotic navigation stacks (e.g., ROS's nav2), where the primary requirement is to know free space versus occupied space.