ETRI’s open Trinity benchmark evaluates planning with 151 user conditions and 95 object types. It checks whether requests such as fetching food respect allergies or dietary needs, distinguishing user-interpretation, environment-exploration and infeasible-action errors. Tested models sometimes understood the environment but chose unsuitable actions. The current food-and-drink virtual setting does not establish performance or safety in real homes or robot hardware.
Back to Tech
ETRI Trinity evaluates user-aware AI task planning
Virtual planning tests do not establish real-world robot safety.