Research

Active Perception

Active perception is grounded in a paradigm shift that computer vision underwent as it moved beyond passive, all-at-once scene analysis. Rather than treating a camera or a model as a fixed device that must take in an entire scene and try to make sense of everything it sees, active perception treats the observer as something that can choose — where to look, how to move through and interact with a scene, or which view to prioritize — in order to answer a specific question as efficiently as possible.

본문

At the lab, this paradigm is most directly instantiated by embodied and robot vision research, where perception is coupled to action rather than passively describing a fully given scene. Affordance-grounding work builds actionable, task-driven representations from partial observations — learning open-vocabulary affordances at scale, and actively reconstructing the unseen 3D structure needed to ground an affordance through generative completion. Compact world models and 3D grounding support the planning and exploration side of the same loop, letting an agent reason about where to move or look next, while extracting relevant visual information rather than processing every view of a scene.

Across these works, the shared commitment is building perception systems that actively reconstruct, ground, and select information in service of action, rather than passively consuming a fully observed input.

Research

본문

Related papers