Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Craftium: Bridging Flexibility and Efficiency for Rich 3D Single- and Multi-Agent Environments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Craftium, a Lua-scriptable 3D voxel environment platform built on a modified Luanti engine that offers Gymnasium/PettingZoo interfaces, runs faster than MineDojo, and supports single- and multi-agent tasks.
Problem
Rich 3D environments for autonomous agent research are computationally expensive (Minecraft-based) and closed-source, while flexible environment tools are mostly 2D or use DSLs; there is a lack of customizable, efficient, multi-agent 3D environments.
Approach
Build the environment platform on the open-source Luanti voxel engine, modified with RL hooks (reward, termination, soft reset, synchronous client/server updates), and expose environment logic through general-purpose Lua mods with standard Gymnasium and PettingZoo interfaces.
Key results
- Craftium reaches 2746.69 steps/s single-environment vs 71.87 for MineDojo (about +2670 steps/s) and 2091.91 for VizDoom.
- Parallel asynchronous Craftium environments exceed 12K steps per second on 16 cores.
- Craftium uses 660MB memory versus 1.7GB for MineDojo.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Craftium instead of MineDojo | Environment framework performance benchmark · Framework performance benchmark | steps/s: +2670 steps per second | Helped |
| Use Craftium instead of VizDoom | Environment framework performance benchmark · Framework performance benchmark | steps/s | Mixed |
Surprising findings
- PPO+LSTM only completes the first open-world stage (collect wood), while zero-shot LLaVA-Agent also unlocks the stone stage.
Limitations
- Maximum number of agents limited to CPU core count
- MineDojo comparison limited to single-environment setups; no parallel support for MineDojo
- Open-world examples are illustrative, not formal benchmarks
Experimental setup
Datasets: Craftium environments (ChopTree, SpidersAttack, Room, SmallRoom, Speleo, Multi-Agent Combat, open-world VoxeLibre, proce
Benchmarks: Craftium environments VizDoom environments MineDojo tasks
Metrics: steps per second episodic return memory usage
Baselines: MineDojo VizDoom Craftax
Builds on: Luanti Gymnasium PettingZoo PPO
Closest papers by meaning
- Craftax: A Lightning-Fast Benchmark for Open-Ended Reinforcement Learning (ICML 2024 · similarity 0.88)
- SimWorld Studio: Automatic Environment Generation with Evolving Coding Agent for Embodied Agent Learning (NeurIPS 2026 · similarity 0.87)
- A Domain-Independent Agent Architecture for Adaptive Operation in Evolving Open Worlds (AAAI 2025 · similarity 0.86)
- CrafterDojo: A Suite of Foundation Models for Building Open-Ended Embodied Agents in Crafter (NeurIPS 2026 · similarity 0.86)
- MCU: An Evaluation Framework for Open-Ended Game Agents (ICML 2025 · similarity 0.86)
- ADAM: An Embodied Causal Agent in Open-World Environments (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.