ROBBRIEF
A new arXiv preprint called FUSE names a long ignored problem in robot perception: recognizing an object by what it does, not by how it looks. The paper also ships a Habitat based benchmark — a standard 3D indoor simulator — for testing it.
A robot asked to find something to sit on in a half-observed room hits a problem most object recognition systems side-step. Most models recognize objects by appearance, and the back of a chair glimpsed from the wrong angle does not read as a chair, let alone as something to sit on. A new arXiv preprint, FUSE, formalizes that gap as its own task and tries to close it by teaching the agent where to look next.
The paper introduces "Active Functional Affordance Grounding," the problem of locating an object that satisfies a functional request, such as "find something to drink from," rather than a named category. FUSE pairs an explicit uncertainty-driven exploration signal with a learned amortized planner, a policy that proposes the next informative viewpoint rather than re-planning from scratch. The authors also release a new Habitat-based benchmark for the task.
On that benchmark, FUSE reports the highest non-oracle grounding performance, the best result among systems without access to ground-truth scene information, and cuts computation by 1.33x relative to fully explicit exploration. "Non-oracle" is a key qualifier: a state-of-the-art claim would require independent benchmark comparisons.
The work is preprint-stage, evaluated on a single simulator, and the gain is incremental over a strong baseline. The contribution is the named problem framing and a new benchmark, not a deployed capability.