BeyondCSe lets a robot grasp "the thing I just used": 77% success on occluded targets, 95% under heavy occlusion with mean views cut from 3.35 to 2.20
The work presents BeyondCSe, a zero-shot robotic grasping system that uses off-the-shelf multimodal large language models and a point tracker to recover a 3D event prior for an event-referenced target and, when the target is occluded, combines that prior with current scene geometry to actively select camera viewpoints; in real-robot experiments with a single wrist-mounted RGB-D camera it reaches 76% and 77% grasp success for initially visible and initially occluded targets (versus 40% and 55% for the strongest baseline), and on four heavily occluded scenes it raises success from 75% to 95% while reducing mean views from 3.35 to 2.20.