Explore high-quality datasets for your AI and machine learning projects.
PARTNR is a benchmark dataset created by FAIR Meta for studying planning and reasoning tasks in human‑robot collaboration. It contains 100,000 natural‑language tasks covering 60 houses and 5,819 unique objects, designed to simulate cooperative scenarios in everyday household activities. The data were generated through a semi‑automated pipeline that combines large language models (LLMs) with a simulated environment, emphasizing constraints on spatial, temporal, and heterogeneous agent capabilities. PARTNR is intended to advance robot‑human collaboration in complex tasks, addressing current model shortcomings in coordination, task tracking, and error recovery.
