Beyond the Current Scene: Event-Referential Grasping with Active View Selection
BeyondSCe enables zero-shot robotic grasping of event-referenced objects, including occluded targets, using pretrained models and active view selection.
BeyondSCe is a zero-shot robotic grasping system for requests that identify an object by its role in a past event, even when the target is no longer visible. It recovers an event prior from interaction history, localizes the object or part, and selects camera viewpoints likely to reveal occluded targets using current scene geometry. With a single wrist-mounted RGB-D camera and no task-specific training, real-robot grasp success is 76% for visible targets and 77% for occluded ones, versus 40% and 55% for the strongest baselines. On four heavily occluded scenes it raises success from 75% to 95% while cutting mean views from 3.35 to 2.20.
- BeyondSCe grasps objects referred to by past interaction roles, not names or appearance.
- Uses pretrained models with no task-specific training and a wrist-mounted RGB-D camera.
- Grasp success is 76% when visible and 77% when initially occluded.
- On heavily occluded scenes, success rises from 75% to 95% with fewer views.
From the original article
1 min read at huggingface.co
A robot that observes people interacting with objects should be able to carry out later requests that refer back to those interactions. Such requests may specify a grasp target by the role it played in a past event rather than by its name or appearance. Moreover, the target may no longer be visible when the robot is asked to act.
By Hyunjoon Lee, Haebeom Jung, Eunsung Cha · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.