Arena 4.0: A Comprehensive ROS2 Development and Benchmarking Platform for Human-centric Navigation Using Generative-Model-based Environment Generation

TL;DR

Arena 4.0 addresses the sim2real gap and platform fragmentation in social navigation benchmarking by delivering a ROS2-based development and benchmarking platform. It combines a generative-world pipeline (Arena-gen) with a semantically linked 3D asset database (Arena-Models) and integrates HuNavSim for realistic human behavior, enabling diverse, production-ready environments. Key contributions include a full ROS2 migration, a two-stage world-generation pipeline driven by LLMs and diffusion models, and a web-enabled usability and benchmarking platform that supports competitions. The framework is validated through a user study showing usability gains and a large-scale evaluation demonstrating scalable world complexity, underscoring its practical impact for rapid development, benchmarking, and deployment on modern robotic platforms.

Abstract

Building on the foundations of our previous work, this paper introduces Arena 4.0, a significant advancement over Arena 3.0, Arena-Bench, Arena 1.0, and Arena 2.0. Arena 4.0 offers three key novel contributions: (1) a generative-model-based world and scenario generation approach that utilizes large language models (LLMs) and diffusion models to dynamically generate complex, human-centric environments from text prompts or 2D floorplans, useful for the development and benchmarking of social navigation strategies; (2) a comprehensive 3D model database, extendable with additional 3D assets that are semantically linked and annotated for dynamic spawning and arrangement within 3D worlds; and (3) a complete migration to ROS 2, enabling compatibility with modern hardware and enhanced functionalities for improved navigation, usability, and easier deployment on real robots. We evaluated the platform's performance through a comprehensive user study, demonstrating significant improvements in usability and efficiency compared to previous versions. Arena 4.0 is openly available at https://github.com/Arena-Rosnav.

Figures (7)

• Figure 1: Arena 4.0 is a software stack of tools to develop and benchmark social navigation algorithms. It fully operates under ROS2 making it compatible with state of the art hardware and eases production level development. We introduce a novel dynamic world generator using generative models, indicated as Arena-gen. Together with various other usability enhancements, Arena 4.0 is highly suitable for unified benchmarking and development. It was selected as the host platform for the https://socialnav2025.pages.dev/organization.
• Figure 2: System Design of Arena 4.0. One of the main contributions is the improved dynamic world generation module, indicated as Arena-gen, using generative models and a language model based interface, which is accompanied by a web-based interface allowing users to customize world generation processes. Closely related, the Arena-Models module, which is a database that allows users to add and semantically link 3D models and assets for the system to plan for and spawn them in a dynamic manner. The Simulation-Setup and Task Generator modules from previous versions have been adapted and extended to take the generated worlds and semantic information as input and generate scenarios and tasks, now additionally based on scene graphs and room segmentations. As already introduced in previous works, our system design is completely abstracted and can thus be operated on multiple different simulators. The whole system has also been migrated to ROS2.
• Figure 3: Data flow of the Generation Stage and the Population Stage. The Generation Stage combines multiple SotA technologies to process text inputs into a floor plan image and room asset locations. 3DSGs are used as an intermediate data structure to divide the problem into a text transformation task solvable by an LLM, and a graph transformation task solvable by a spatial GNN. The Population Stage populates the floor plan's asset zones with 3D models by employing the Asset Placer. A pre-built semantic vector Model Database is queried for a related model, which is arranged into the zone by a Fitter algorithm. After a final post-processing step, the end result is a finished environment consisting of 3D walls and models.
• Figure 4: GUI of the Model database. The user is able to integrate their own 3D objects, annotate them, link them semantically with rooms, other objects, or interactions with humans. Objects and assets are queriable by a number of attributes such as color, type, probability to appear in a specific room, their annotations and links to other entities such as rooms or other objects. Using the GUI, the user can specifically set the attributes or include new objects easily.
• Figure 5: Example worlds generated using the Arena-gen module for benchmarking and competition purposes. The worlds were generated using the text "generate me 5 difficulty levels of a [hospital, residential, office] environment". The assets are automatically taken from the arena model database and pedestrians spawned with HuNavSim. Notably, a large variety of worlds for each environment type and level can be generated, e.g. 500 environments of hospital level 2. This feature aids quantitative benchmarking and in training new models. Users can also customize room layouts, pedestrian interactions with each other, asset placements, and specific situations using the Arena Architect GUI (shown in the supplementary video).