meta-learning

We introduce a novel meta-learning grasp framework aimed at addressing the relatively unexplored challenge of grasping objects characterized by diverse physical properties, relying solely on visual input. Two innovative synthetic datasets that explicitly incorporate physical properties are introduced, making them compatible with a wide range of simulation frameworks. Our approach demonstrates substantial advantages in real-world object manipulation, despite being trained exclusively in a simulated environment.

This work investigates meta-learning algorithms on vision regression tasks and demonstrates their ability to tackle structured problems with a new proposed functional contrastive learning on the task representation of CNPs to improve its expressivity. Furthermore, it quantitatively analyzes various deep learning techniques to alleviate meta overfitting.

We introduce a novel meta-learning framework for 6D pose estimation with strong generalization ability on unseen objects within and across object categories, with a proposed GNN-based keypoint prediction module that leverages geometric information from canonical keypoint coordinates and captures local spatial constraints among keypoints via message passing.