Illustration brief the model wrote
A 3D scene graph hovering above a robotic wheeled platform in a warehouse aisle; the graph has nodes for objects with 'Recall', 'Explore', and 'Answer' arrows. The robot's camera is turned toward a partially opened door, with a glowing question mark above it.
Enter the Mind Palace: Long-term Active Embodied Question Answering
Defines long-term active embodied QA with scene-graph memory and value-of-information stopping for autonomous recall, explore, or answer decisions.
What changed
ModelPrior embodied QA systems operated in short sessions with no persistent memory, forcing agents to re-explore environments. This paper introduces long-term memory (a scene-graph 'mind palace') and a value-of-information policy that lets the agent decide when to recall past observations, explore new areas, or directly answer a question across days or weeks.
So what, in plain English
ModelImagine a detective who can build a mental map of a building over several days, then when asked a question decide: do I remember the answer from before, do I need to search a new room, or can I answer now? This paper gives robots that same ability to gather information over long periods and intelligently choose what to do next.
Contribution to industry
ModelThis paper provides a memory architecture and decision policy that can be integrated into inspection and patrol robots for industrial facilities, enabling them to answer questions about changes over days or weeks without constant human guidance. It contributes a formal definition of the task, a benchmark environment, and a method that balances exploration and memory retrieval—directly useful for autonomous facility monitoring and security robots.
Typed judgments
Jev★★☆☆☆ embodied · Physical AI 93%
Real hardware 26% · Open code / data 9% · Foundation-model based 41% · Survey / benchmark-only 4%
Probabilities from TypeSafe Jev over the title, venue and abstract; the filters on the atlas apply code-owned thresholds to these numbers.
Editor's notes
Summary. Defines long-term active embodied question answering and solves it with a scene-graph "mind palace" memory plus value-of-information stopping, so the agent decides when to recall, when to explore and when to answer.
Industry. Gives inspection and field robots a persistent, queryable memory of sites they patrol over days and weeks — a requirement for industrial facility autonomy.