AWhat changed
“Is there an apple in the refrigerator?” cannot always be answered from the first camera image. The agent may need to move, open the door and remember what it saw. Gordon and colleagues introduce an interactive question-answering task and a hierarchical system that divides planning, navigation, detection, manipulation and answering. A spatial memory stores observations at locations rather than forcing a single recurrent state to remember everything. Their IQADATA resource supplies automatically generated existence, counting and spatial-relation questions in simulated kitchens. The contribution combines this task design with the specific memory and controller architecture, without separately counting the underlying AI2-THOR engine.
Scientific scope
“Is there an apple in the refrigerator?” cannot always be answered from the first camera image. The agent may need to move, open the door and remember what it saw. Gordon and colleagues introduce an interactive question-answering task and a hierarchical system that divides planning, navigation, detection, manipulation and answering. A spatial memory stores observations at locations rather than forcing a single recurrent state to remember everything. Their IQADATA resource supplies automatically generated existence, counting and spatial-relation questions in simulated kitchens. The contribution combines this task design with the specific memory and controller architecture, without separately counting the underlying AI2-THOR engine. The original split uses 25 training kitchens and five test kitchens; questions follow templates rather than unrestricted human conversation. Training uses simulator-provided supervision, and some reported results additionally use ideal detections or navigation. Those results must be distinguished from learned-perception results. The learned system does not outperform simple answer-frequency guessing on every question type. No real-house robot experiment or general language-understanding guarantee is established.
Evidence
The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The original split uses 25 training kitchens and five test kitchens; questions follow templates rather than unrestricted human conversation. Training uses simulator-provided supervision, and some reported results additionally use ideal detections or navigation. Those results must be distinguished from learned-perception results. The learned system does not outperform simple answer-frequency guessing on every question type. No real-house robot experiment or general language-understanding guarantee is established.
Contributors
- The named authors in the research record · SCIENCE: Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.
Provisional WRS assessment, edition 2026-10-23.6. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction