# Goal-Directed Embodied Exploration

The authors introduce RoboQuest, a benchmark for goal-directed embodied exploration with ten mobile manipulation tasks spanning search, manipulation-based inspection, and interactive testing; five frontier multimodal agents run zero-shot through a common visuomotor interface and the best succeeds in only 23.2% of episodes, while a VLA policy fine-tuned on the released 5,000 demonstrations almost never succeeds; isolated skill tests show agents can perform most required actions when hidden information is supplied, and failure analysis attributes only 18–21% of failures to execution, with most arising from stopping exploration too early and deciding before observing required evidence.

The authors introduce RoboQuest, a benchmark for goal-directed embodied exploration with ten mobile manipulation tasks spanning search, manipulation-based inspection, and interactive testing; five frontier multimodal agents run zero-shot through a common visuomotor interface and the best succeeds in only 23.2% of episodes, while a VLA policy fine-tuned on the released 5,000 demonstrations almost never succeeds; isolated skill tests show agents can perform most required actions when hidden information is supplied, and failure analysis attributes only 18–21% of failures to execution, with most arising from stopping exploration too early and deciding before observing required evidence.