BesiegeField – LLM Agents Learn to Build Machines in a Physics Sandbox
BesiegeField – LLM Agents Learn to Build Machines in a Physics Sandbox
Hi HN! We'd like to share something we built: BesiegeField — an environment where large language models design, test, and refine machines in the physics-based construction game Besiege. Machine design is treated as a code generation problem: LLMs choose standard parts and specify how to connect them, making the task well-suited to language models. The environment includes tasks like throwing a stone far or navigating bumpy terrain, and supports custom goals and environments. We used it to explore both agentic workflows (no finetuning) and reinforcement learning, trying tasks like building cars and catapults. It runs 100+ parallel processes on Linux clusters for scalable RL training. Feedback from the AI, RL, engineering, and game dev communities is welcome. GitHub: https://github.com/Godheritage/BesiegeField HuggingFace Demo (single-agent): https://huggingface.co/spaces/Godheritage/BesiegeField-Machi...
Share cardActual performance
Launch Intel predictions
Analyze your own launch →Incorrect prediction on native model
Similar products
Phynix = Sandbox + Physics + Unix
TypeGPU Realtime Physics Sandbox
Physics Sandbox Game Using WASM a.k.a. JollyWorld
Phynix, a Physics Sandbox with Unix-like Commands
A Mindmap of Physics
FOSS physics-based sandbox game with circuiting, robotics, scripting
Azure-sandbox
Helios OpenTelemetry Sandbox
ECMAScript Sandbox
Spacedeck Sandbox