HuggingFace Daily PapersZecong Tang
HiSTEVE 1.0: Harnessing Human-Interface VLA Agents for Open-World Games
Papers83
HiSTEVE 1.0 presents a hierarchical Vision-Language-Action (VLA) framework that navigates Minecraft strictly through human-native keyboard and mouse inputs, bypassing privileged game APIs. A high-level VLM orchestrator decomposes tasks using structured memory, while a single unified VLA executes motor control with dedicated RMSNorm branches separating GUI menus from 3D camera rotation. Scoring 85.0% on MineStudio Bench and 37.4% on OpenHA, the system successfully tackles long-horizon challenges such as crafting an iron pickaxe from scratch.
Why it's worth reading
It demonstrates how open-world embodied agents can bridge high-level planning and fine-grained motor control purely through human-like interfaces, bypassing privileged engine hooks.
Tags
Embodied AIVLAMinecraftReinforcement LearningAutonomous AgentsMultimodal