HiSTEVE 1.0: Harnessing Human-Interface VLA Agents for Open-World Games

One-line takeaway

HiSTEVE 1.0 achieves state-of-the-art atomic task performance and long-horizon gameplay in Minecraft via a hierarchical VLA agent with structured planning and native interface control.

HiSTEVE 1.0 提出一种层级式视觉语言代理，结合结构化记忆与规划模块，在Minecraft中实现了长时域自主决策和原生界面控制，达到顶尖的原子任务表现并完成从零开始的复杂建造任务。

Key points

HiSTEVE 1.0 introduces a hierarchical visual-language agent for Minecraft that integrates long-horizon planning with native GUI control. It uses structured memory, a VLM planner, and policy-coupled completion detection to manage subtasks. Separate regime-specific branches and embeddings mitigate interaction interference. The agent achieves top-tier performance on benchmark tasks and successfully completes complex, human-like long-horizon gameplay.

• Proposes a hierarchical VLA agent with structured memory and VLM planning for open-world autonomy.
• Introduces policy-coupled task completion detection to reliably signal subtask end based on agent policy context.
• Uses regime-specific RMSNorm branches and separate mouse embeddings to reduce GUI-non-GUI interaction interference.
• Achieves 85.0% atomic task success on MineStudio Bench and completes long-horizon crafting tasks like human players.

Written only from the official arXiv abstract by an open-source model, then checked automatically (every number and name must appear in the abstract). The paper is authoritative.