ROBBRIEF
A new preprint argues the bottleneck is not what the robot sees but the order in which it thinks about the objects, using a language model's commonsense about kitchens to rank what matters.
A home robot asked to fetch a mug from an unfamiliar apartment can map the room and spot the counter, but it has to decide which of the dozens of objects it sees deserve attention first, and which are filler. A new preprint on object-goal navigation argues the bottleneck is not perception but that ordering problem.
Most prior work treats objects as a bag, in any order. The proposed Graph-MambaNav method instead gives the robot a heuristic ranking: items more likely to relate to the target are processed later in a spatial-temporal graph scan. Later nodes get to summarize what earlier ones passed along, so the model weighs a fridge higher than a wall clock when the target is a mug.
The ranking and the edges between objects are seeded by a large language model's commonsense about which objects tend to co-occur in which rooms, instead of learning those priors from pixels alone. A spatial module mixes local message passing with a global selective scan across the ordered graph; a temporal module then chains those scans over time so the robot can carry its conclusions between steps.
The authors report gains on the AI2-THOR and RoboTHOR simulators, plus a real-world robot deployment as a lab validation, not a product milestone. The paper is a preprint, not peer reviewed, and the gains are incremental. The ordering trick, though, is portable: when an agent faces a crowd of candidates, processing the most informative ones last is something language models have done for years, and now a home robot can borrow it.