Inside MapScale® Galileo: Agentic AI and AI-Native Mapping

A few months ago, we shared a major milestone with MapScale® v9.0, achieving a +34% improvement in unit classification accuracy by integrating multi-modal LLMs.

This enabled MapScale® to work with a much broader range of building data, including scanned floor plans, rasterized drawings, handwritten annotations, embedded text, and other unstructured inputs that are commonly found in real-world projects.

Building on this foundation, our latest model, MapScale® Galileo, further improves the indoor mapping pipeline through three core technical components:

• Multi-modal LLMs for simultaneous visual and textual data interpretation.
• An agentic AI framework that coordinates specialized sub-proceszes for floor plan parsing
• A simplified, structured taxonomy designed for AI mapping
  

What’s the Major Breakthrough?

Building on the multi-modal LLM foundation introduced in MapScale v9.2, Galileo adds two key components: a simplified, AI-native taxonomy and an agentic AI approach.

The taxonomy provides a consistent structure for representing indoor spaces, while the agentic framework coordinates specialized mapping tasks such as space classification, metadata extraction, and validation.

Individually, each component improves map generation accuracy and consistency. Combined, they enable more reliable interpretation of complex floor plans and produce higher-quality indoor maps with less manual intervention.

A Simpler, AI-Native Taxonomy

One of the biggest challenges in AI mapping is space classification. Many widely adopted mapping models, such as Google Maps, Apple IMDF, and OpenStreetMap, are designed for outdoor mapping, wayfinding, or data exchange. While somewhat effective for their intended use cases, these models can introduce additional complexity when applied to AI-driven indoor mapping, where spaces, amenities, accessibility features, and points of interest need to be classified consistently across different buildings.

Comparing Google Maps, Apple IMDF, OpenStreetMap, and Pointr Maps

To illustrate the differences, consider how different models handle a specific real-world example: A wheelchair-accessible Mexican restaurant with a vegan menu.

Google Maps (Simple but limited hierarchy)

"primaryType": "restaurant",

"types": [
    "mexican_restaurant", 
    "vegan_restaurant", "food"
          
],

"accessibilityOptions": {

    "wheelchairAccessibleRestroom": true

}

Advantage

Effective for broad, outdoor-focused categorization. The flat structure provides a standard entry point for general place identification.

Limitations

Lacks the hierarchy necessary for complex indoor environments. Contextual data is fragmented across multiple fields, limiting extensibility for nuanced spatial classification.

Apple IMDF (Detailed but complex)

"feature_type": "occupant", 

"properties": {  

    "category": "mexican.modern",  

    "anchor_id": "123" 

}

Advantage

Detailed and structured. The rich category model supports precise representation of indoor spaces and features.

Limitations

Complex structure. The use of 900+ predefined categories and multiple linked entities can make classification harder to automate and maintain consistently at scale.

OpenStreetMap (Flexible but inconsistent)

"amenity": "restaurant"

"cuisine": "mexican"

"diet:vegan": "yes"

"wheelchair": "yes"

Advantage

Flexible and community-driven. Contributors can add detailed information without being limited to a fixed set of categories.

Limitations

No formal hierarchy. Because structure depends on individual tags, similar spaces can be modeled differently (e.g., amenity vs. shop), making consistent AI-driven classification and validation more difficult.

The Pointr Maps Approach (Structured and extensible)

Pointr Maps introduces a clean, two-level taxonomy:

"mainType": "food-beverage",

    "subType": "restaurant",

      "cuisine": "Mexican",

      "dietaryOptions": ["Vegan"],

      "wheelchairAccessible": true

Key Principles:

• mainType (required): core category
• subType (optional): adds detail when needed
• Attributes separated from taxonomy (optional): flexible and extensible

Instead of relying on complex hierarchies or inconsistent tagging, it separates core structure (mainType, subType) from flexible attributes like "cuisine" or "wheelchairAccessible".

This makes it easier to configure, faster to scale, and more reliable for AI to classify and identify spaces, resulting in higher-quality maps with less manual effort.

The New Frontier: Agentic AI Mapping

Once spaces can be represented consistently, the next challenge is how floor plans are interpreted and converted into map data.

In our previous AI mapping models, we followed a fixed pipeline, where one model detects geometry, another classifies spaces, and another extracts metadata. Each step runs once in sequence. While effective, this approach has limitations: errors can propagate through the pipeline, earlier decisions cannot be revisited, and all floor plans are treated the same regardless of complexity.

MapScale® Galileo introduces an agentic approach by combining multiple specialized models that collaborate during the digitization process to interpret floor plans. Instead of following a fixed sequence, the system breaks the problem into smaller tasks, chooses how to solve each step, and allows models to validate and refine each other’s outputs.

Galileo also analyzes floor plans at multiple scales and resolutions to capture both high-level structure and finer details. It can allocate additional processing to complex regions and leverage CAD metadata, such as layers, names, and annotations, to improve contextual understanding. Outputs from multiple specialized models are then combined into a single, consistent map representation, improving accuracy while reducing manual mapping effort.

Benchmark Results

We evaluated Galileo against previous MapScale® versions.

Key results:

• Detection remains industry-leading at 90.2% (identifying spatial geometry (walls, boundaries, POIs)
• Classification reaches up to 89.4% (understanding what a space is, e.g., patient room vs pharmacy)
• Identification improves to 81% (extracting precise details, e.g., Room 101 vs Room 102)

Compared to MapScale v9.2, this represents ~10% improvement in both classification and identification.

What’s Next?

Galileo is a major step forward, but it’s part of a broader shift.

We’re moving toward:

• Fully automated map generation
• AI-ready indoor data for agentic systems (MCP)
• Standardized, scalable indoor mapping across industries

With continued improvements in taxonomy, data coverage, and AI models, MapScale® is evolving into the infrastructure layer for indoor spatial intelligence.

Contributor