Awesome AI Agents › Research Platforms & Simulators

SimWorld-AI/SimWorld

⭐ 779 Python repository created 2025-04-07

SimWorld is a simulation platform built on Unreal Engine 5 for developing and evaluating autonomous agents driven by large language and vision-language models. It provides open-ended world simulation with realistic physics and language-based procedural generation, so scenes and city layouts can be produced from natural language rather than authored by hand. Agents connect through a Python interface that exposes multi-modal perception and natural-language actions, and the quick-tour example follows a gym-like loop familiar from reinforcement learning environments. The project is organized in three levels of use. The base package includes two lightweight city scenes plus an empty map for core interaction and quick testing. An optional environment pack adds more than a hundred pre-built maps for richer scenarios. Full customization lets users bring their own Unreal environments, assets and agent models, with documentation covering how to package custom content. The stated research focus is physical and social reasoning: navigation, planning, multi-agent coordination and other complex behaviours that need an embodied setting rather than a text sandbox. The repository ships the Python client, the Unreal server, configuration files, an API reference and read-the-docs documentation, along with demo videos. The work has an arXiv white paper, was accepted as a spotlight at NeurIPS 2025 and appeared in the CVPR 2025 demonstration track. It suits researchers who study embodied or socially situated agents, robotics-adjacent navigation work, and anyone who needs a controllable open world in which many agents act at once.

https://github.com/SimWorld-AI/SimWorld

simulatorunreal-engineembodied-agentsllm-agentsmulti-agentnavigationresearchvlm