ICRA 2026 Workshop  ·  Robots Meet Prior Maps

RoboBIM: Scaling Robot Semantic Foundations through Agentic Extraction of BIM Priors

Ashwin Nedungadi*, Bharathi Kannan Nithyanantham*, Stefan Luedtke

Institute for Visual & Analytical Computing, Universität Rostock   ·   * Equal Contribution

Code Paper (PDF)
RoboBIM hero: multi-layer floorplan navigation graph
Abstract

Building Information Models (BIM) stored in the Industry Foundation Classes (IFC) format contain rich priors for indoor robot autonomy. However, most pipelines use static one-pass conversions that produce occupancy maps or simulation assets while leaving much of the semantic information unused.


We present RoboBIM, an agentic pipeline that converts IFC files into metric-semantic navigation contracts. RoboBIM separates structural extraction from semantic enrichment. First, a deterministic backbone creates a reliable geometric baseline including occupancy maps, waypoints, and room topology. Then, an LLM agent infers missing semantic information — room types, hazards, and traversability — through sandboxed code execution. We also introduce an online query agent that retrieves these priors during robot operation.

TL;DR

RoboBIM converts IFC building models into semantically rich robot navigation contracts using a deterministic graph backbone + a 7-stage LLM enrichment layer, increasing room-type coverage by 74.5% and filling hazard, material, traversability, and policy fields left empty by the deterministic baseline.

+74.5%
Room-type coverage increase
4
Public IFC models evaluated
7
Agentic enrichment stages
131
Rooms processed (51→89 typed)
3
Export formats (Nav2, ROS2, GraphML)
403 s
Mean wall-clock per building
Contributions
What RoboBIM Introduces
🗺️
Three-Layer Navigation Graph
A per-storey occupancy grid, clearance-aware waypoint graph, and topological room-transition graph extracted from IFC in a top-down manner. Exports directly to Nav2, Open-RMF, GraphML, and GeoJSON.
🤖
Staged LLM Agent Pipeline
A seven-stage agent that tunes backbone hyperparameters within fixed ranges, infers missing semantic fields, and produces a deterministic natural-language briefing for an onboard robot LLM.
📊
Preliminary Evaluation on IFC Bench
Evaluation on four public IFC models shows RoboBIM increases room-type coverage, fills hazard, material, traversability, and policy fields that are null in the static baseline.
Method
System Architecture

RoboBIM has two main parts: a deterministic geometric backbone and an LLM agent wrapper. The backbone first extracts the structural graph Gstatic = (𝒪, 𝒲, 𝒯) and then the seven-stage agentic layer enriches this graph to produce Gagent = (𝒪, 𝒲, 𝒯, Σ).

RoboBIM Pipeline
🏗️

System Pipeline Figure
Static backbone (left) vs. RoboBIM agent pipeline (right).
See paper Fig. 2 for full detail.

Fig. 2: Static pipeline vs. RoboBIM agent pipeline. The static pipeline is five deterministic modules that produce Gstatic. The agent pipeline wraps the same backbone with seven LLM stages and a topology regression gate, producing Gagent.
Output
Three-Layer Navigation Graph

The enriched graph 𝒢agent has three linked layers. Each room is linked to its waypoint cluster, and each transition is linked to its crossing waypoint, supporting hierarchical planning.

Layer 1: Occupancy Grid
Layer 1 · Metric
Occupancy Grid
Per-storey occupancy grid with 5 cm resolution. Each cell is labelled as free, occupied, or unknown. Forms the metric foundation for trajectory planning.
Layer 2: Waypoint Graph
Layer 2 · Affordance
Waypoint Graph
Per-storey waypoint graph derived from the GVD skeleton. Each edge stores a clearance value c ∈ ℝ≥0 determining whether a robot with inflated radius r can traverse the edge.
Layer 3: Topological Graph
Layer 3 · Semantic
Room Topology
Topological graph of rooms and transitions including doors, wall openings, and stair/elevator landings. Enriched with room types, hazard codes, and traversability scores.
Agentic Enrichment
Seven-Stage LLM Pipeline

The agent wrapper adds seven stages. Each stage reads from and writes to a shared Pipeline Context. A topology regression gate audits every edit and rolls back any change that reduces the connected-transition count below the deterministic baseline.

0
IFC Exploration
Sandboxed Python inspection of the IFC file. Lists entity counts, storeys, doors, and representative property sets.
1
Quality & Parameter Tuning
Returns a completeness score and five numeric parameters: grid resolution, dilation radius, min room area, door search radius, include-openings flag.
2
Room-Type Inference
Regex and property-set prefilter handles obvious names. Remaining rooms on each storey sent to the LLM in a single batched prompt. Confidence threshold ≥ 0.80.
2b
Segmentation Review
Reviews the room list and transitions; proposes merge or split directives. Merges with confidence ≥ 0.70 are applied; splits are recorded but not applied.
3
Connectivity Validation & Repair
Produces a repair plan with add_connection, remove_connection, reclassify directives. Topology regression gate audits the connected-transition count.
4
Hazards & Traversability
Writes hazard codes {wet, narrow passage, fall risk, high traffic, obstacle dense, restricted} and traversability score in [0, 1]. Floor material strings added.
5
Post-Review & Robot Briefing
Final natural-language audit and briefing designed to be copied directly into a robot LLM system prompt. Robot needs no runtime IFC access.
Experiments
Results on IFC Bench

We evaluate RoboBIM on four public IFC models from IFC Bench: 4351, BasicHouse, digital_hub, and samuel_macalister_sample_house. These cover residential, mixed-use, and office layouts with 14–64 rooms across 2–6 storeys.

Model Storeys Geometry (preserved) Semantic Enrichment (static → agent) Agent Cost
RoomsConn.Waypoints Typew/ HazardTraversabilityMaterial Time (s)Tokens in/out
435152290— 1 → 30 → 220 → 50 → 5 294285k/32k
BasicHouse214632 0 → 70 → 70 → 140 → 7 489200k/24k
digital_hub36460304 38 → 640 → 00 → 640 → 1 423343k/32k
macalister6319392 12 → 170 → 90 → 310 → 6 407348k/41k
Total1613185818 51 → 890 → 190 → 1310 → 18 16121176k/129k

Geometry columns are identical for both pipelines (gate forbids regression). Semantic enrichment columns: static → agent.

Per-stage cost breakdown
Stage cost breakdown. Agent wall-clock time ranges from 294 to 489 s per IFC model (mean 403 s). Stage 2b is the main cost on BasicHouse; Stage 3 is the main cost on digital_hub. Per-storey batching keeps LLM calls between 7 and 15 per run.
Citation
Cite This Work

If you use RoboBIM in your research, please cite our workshop paper: 

@inproceedings{nedungadi2026robobim,
  title     = {{RoboBIM}: Scaling Robot Semantic Foundations
               through Agentic Extraction of {BIM} Priors},
  author    = {Nedungadi, Ashwin and
               Nithyanantham, Bharathi Kannan and
               Luedtke, Stefan},
  booktitle = {ICRA 2026 Workshop on Robots Meet Prior Maps:
               Leveraging Geometric and Semantic Priors for
               Real-World Robot Autonomy},
  year      = {2026},
  note      = {Equal contribution: Nedungadi, Nithyanantham}
}