FUSE robot framework finds objects by function, cutting compute 1.33x

Embodied agents typically struggle when asked to find an object by its function—like "something to pound with"—when visual cues are occluded. Existing affordance grounding methods assume fixed viewpoints, so they can't reason about where to look. To fix this, Zhou Chen and Sathyanarayanan N. Aakur (arXiv:2608.12683) define the task of Active Functional Affordance Grounding: an agent must sequentially explore a scene, decide where to move, and spatially ground the object that satisfies a functional query. Their proposed framework, FUSE, couples an explicit uncertainty-driven explorer with a learned amortized planner, enabling efficient viewpoint selection without exhaustive exploration.

FUSE was tested on a new Habitat-based benchmark built specifically for active functional grounding. Results show it achieves the highest non-oracle grounding accuracy among baselines while using 1.33x less computation than fully explicit exploration, and it remains robust across multiple affordance knowledge sources. The work is under review, with 15 pages and 9 tables. This research pushes robots closer to practical reasoning about object purpose in real-world environments, particularly for service robots and warehouse automation.