Ask When It Pays: Cost-Aware Open-Ended Interaction for Instance Goal Navigation
Researchers have developed a new approach for instance goal navigation, where an embodied agent must find a specific object based on a vague description. The system treats interaction with an oracle as a cost-sensitive uncertainty-reduction problem, prompting the agent to ask questions that most efficiently resolve ambiguity. This work introduces a new benchmark and a Weighted Success Rate metric to evaluate interaction efficiency, moving beyond previous methods that did not account for varying question costs. AI
IMPACT This research could lead to more efficient embodied AI agents capable of understanding and acting on ambiguous instructions in complex environments.