World Tracer: An Easy-to-use Program that Creates the Layout of 3D Environments from LiDAR SLAM Data

WorldTracer

Kim, You-Jin & Sharkey, Tommy (2023). WorldTracer (Version 1.4.0) [Computer software] Repository.

Reconstructing 3D factory floor models from live robotic SLAM data for DNN agent training.

▪️  DOI: 10.5281/zenodo.6565136    ▪️  GitHub: github.com/yujnkm/WorldTracer 

Abstract

World Tracer is a rapid 3D environment generator designed to bridge the gap between 2D SLAM mapping and high-fidelity physics simulations by providing accurate, scale-corrected, and real-time layout updates. By utilizing bitmap data from robot LiDAR and SLAM sensors as a structural scaffold, this project enables the near-instantaneous transformation of 2D floor plans into 3D digital twins through a streamlined tracing and extrusion process. The system facilitates a live data pipeline that generates updated training environments by retrieving optimized OBJ files directly from the server after every scan. Because these models are lightweight and include complete 3D meshes with rigid-body wall models, they can be processed directly on the server to efficiently analyze ideal parameters and verify spatial constraints before physical implementation.

This automated layer allows the optimization of Deep Neural Network navigation models via Reinforcement Learning within specific areas, focusing on recently changed zones where collisions occur more frequently. During deployment in high-density industrial scenarios involving fleets of up to 70 Omron AMRs featuring five distinct types with unique movement capabilities, agents were trained within the dynamic digital twin to optimize complex kinematic parameters such as linear speed, rotational velocity, and cornering efficiency. This approach allows the system to simulate an exhaustive range of hypothetical conditions, ensuring that navigation policies remain robust even as industrial equipment is relocated or repurposed. By utilizing lightweight models that run directly on the server, the framework maximizes throughput and minimizes traffic congestion without waiting for actual collision data, representing a significant advancement in the scalability and safety of autonomous fleet management.

Research Contributions

• The system automates 2D SLAM and LiDAR conversion, providing updated, lightweight vertex meshes digital twins ready for training. 

• The framework simulates up to 70 Omron AMRs simultaneously to analyze emergent behaviors like traffic congestion and high-density bottlenecks in the map.

• The platform captures real-time environmental states, ensuring agents train on current layouts rather than outdated blueprints to minimize collision penalties.

• This research contributes to spatial computing by procedurally extruding 2D scaffolds into high-fidelity 3D twins without photogrammetry overhead.

• The work establishes a continuous learning pipeline where simulation evolves with the physical world, creating a closed-loop system for fleet optimization.

Citation IEEE Format

[1] Y-J. Kim, "World Tracer," version 1.4.0, Zenodo, 2023. [Online]. Available: https://doi.org/10.5281/zenodo.6565136. GitHub: https://github.com/yujnkm/WorldTracer

Citation APA Format

Kim, Y-J. (2023). World Tracer (Version 1.4.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.6565136; GitHub: https://github.com/yujnkm/WorldTracer