Recent advances in foundation models have spurred efforts to embed agents in the physical world. Such agents must possess spatial agentic intelligence: exploring unfamiliar environments, continuously updating spatial understanding through interaction, and adapting actions based on feedback to achieve a sequence of goals. Existing benchmarks are limited to narrow layout varieties, fixed scales, or simple traversal tasks, which fails to capture the full complexity of real‑world spatial reasoning.
Mine Odyssey is a benchmark built on Minecraft reconstructions of real‑world locations, designed to evaluate agents' spatial agentic capabilities comprehensively. It comprises 180 tasks across 30 locations spanning 20 countries and regions on five continents, including 20 outdoor and 10 indoor settings. The sites range from Midtown Manhattan and rural Entrup to Santa Lucía Hill and Buckingham Palace, covering diverse spatial scales, layouts, terrains, and connectivity patterns.
For each task, meaningful waypoints such as landmarks, buildings, and rooms are selected and manually verified for accessibility. Natural‑language instructions specify the order in which these waypoints must be visited. Success requires agents to discover traversable routes, locate and open doors, transition between levels via stairs or ladders, monitor progress, and recover from navigation errors.
Evaluation of eight state‑of‑the‑art models shows GPT‑6 Astra achieving the highest success rate of 85.6%. The runner‑up Claude Opus 5.5 completes 73.9% of tasks, while the strongest open‑weight model, DeepSeek‑V4.1‑Flash, reaches only 23.9%, highlighting substantial room for improvement in current spatial agentic intelligence. Comprehensive analyses and ablation studies expose limitations in path planning, hierarchical navigation, and error recovery, offering clear guidance for future model development.
Review