Arena-Rosnav: Extensible Navigation Platform

• Arena-Rosnav is an open-source, extensible platform designed for developing, benchmarking, and deploying robot navigation systems in dynamic, human-centric environments.
• It features modular simulation backends—including Flatland, Gazebo, and Unity—and a standardized API that supports both ROS and ROS 2 for seamless planner integration.
• The platform incorporates advanced pedestrian and social force modeling with real-time dynamic obstacle handling, enabling robust sim-to-real performance.

Arena-Rosnav is an open-source, extensible research platform and software suite for developing, benchmarking, and deploying robot navigation systems in highly dynamic and human-centric environments. Evolving across multiple generations, Arena-Rosnav integrates modular simulation backends, sophisticated pedestrian and crowd modeling, classic and learning-based planners, advanced benchmarking metrics, and tooling for seamless sim-to-real deployment. The system is designed for compatibility with both ROS and ROS 2, enabling rigorous cross-comparison of planning approaches in scenarios characterized by dense, realistic human crowds and dynamic obstacles (Kästner et al., 2021, Kästner et al., 2023, Kästner et al., 2022, Kästner et al., 2024, Shcherbyna1 et al., 2024).

1. Architecture and Modular Pipeline

Arena-Rosnav is organized as a layered, plugin-oriented framework built atop the Robot Operating System (ROS/ROS 2). The core system exposes unified APIs for simulators, planners, and evaluators, allowing quick integration of new planning algorithms, simulation engines, and metrics modules.

• Simulation Backends: Includes high-speed 2D (Flatland, Arena2D), physics-based 3D (Gazebo), and photorealistic 3D (Unity) environments. World and scenario generation is mediated via CLI and GUI tools.
• Task Management: A central Task Manager orchestrates robot, pedestrian, and scenario instantiation using YAML/JSON specification files.
• Planning Suite: Supports conventional (A*, DWA, TEB, MPC), hybrid (APPLR, Dragon, TRAIL), and deep reinforcement learning (CADRL, RLCA, ARENA, ROSNavRL) planners. All planners implement a standardized interface for runtime switching and consistent evaluation.
• Pedestrian and Social Force Modeling: Realistic pedestrian dynamics via Pedsim (social-force models), graph-based navigation, and extensions for group/interaction behaviors. Dynamic obstacles publish 10 Hz pose data for real-time navigation.
• API Abstraction (Arena-Core): All critical functions—map/world generation, agent lifecycle, scenario resets, and benchmarking—are exposed via standardized Python/C++ methods and ROS launch configurations. The interface enables multi-simulator experiments with identical task definitions (Kästner et al., 2023, Kästner et al., 2024).

2. Dynamic Obstacle and Social Force Simulation

Human and obstacle modeling in Arena-Rosnav is grounded in multiple parameterized social force models:

• Helbing–Mollnar Classical SFM: For each pedestrian $i$,

$$m_i \frac{d \mathbf{v}_i}{dt} = m_i \frac{\mathbf{v}_i^0 - \mathbf{v}_i}{\tau} + \sum_{j \neq i} \mathbf{f}_{ij} + \sum_{W} \mathbf{f}_{iW}$$

with goal force, pedestrian repulsion, and obstacle repulsion.

• Group Extensions (Moussaïd et al.): Incorporates gaze alignment, group attraction, and intra-group repulsion for modeling complex group behaviors and formations (Kästner et al., 2024).
• Human–Robot Interactions: Additional avoidance forces model pedestrian reactions to robots and permit planners to account for social comfort via semantic layer costmaps.

Pedestrian start/goal assignments may be randomized, scenario-defined, or optimized via cost-based waypoint selection. Pedsim and graph-based controllers are provided, both with stochastic