Researchers have introduced HiSTEVE 1.0, a novel vision-language model (VLA) agent designed to operate within open-ended games like Minecraft. This agent features a hierarchical structure with a capability-aware harness that uses structured memory and a VLM planner for task selection. HiSTEVE 1.0 demonstrates state-of-the-art performance on atomic tasks, achieving 85.0% on the MineStudio Bench and 37.4% on OpenHA, while also successfully completing complex, long-horizon tasks such as crafting an iron pickaxe from scratch. AI
IMPACT This research advances embodied AI capabilities, potentially leading to more sophisticated virtual agents for gaming and simulation.
RANK_REASON This is a research paper detailing a new model and benchmark. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →