1341-UR Active Viewpoint Selection for Robotic Manipulation under Occlusion
We study active perception for a mobile manipulator with a wheeled base, an arm, and an RGB-D camera. A point-and-relation command identifies the manipulation task. Reference and recurrent visual memory guide viewpoint selection, readiness, and contact-pixel prediction for fixed grasp, push, or pull controllers. Training uses supervised pretraining and DAgger. In MuJoCo tests, the full policy achieves 52.7% success under partial occlusion and 42.2% under heavy occlusion.
Authors
- Wooyoung Yun
- Ethan Leitner
Publication Details
- Journal
- DigitalCommons - Kennesaw State University (Kennesaw State University)
- Published
- 2026-11-20
- Primary Topic
- Robot Manipulation and Learning
- Type
- article
- Field-Weighted Citation Impact
- 0.00