Core Concepts
CenterArt is a novel approach for simultaneous 3D shape reconstruction and 6-DoF grasp estimation of articulated objects from RGB-D images.
Abstract
CenterArt is a vision-based approach that addresses the challenge of manipulating articulated objects. It consists of an image encoder that predicts object heatmaps, poses, shape codes, and joint codes, and a decoder that reconstructs 3D shapes and estimates valid 6-DoF grasp poses.
The key highlights of CenterArt are:
It is the first approach that can simultaneously perform 3D shape reconstruction and 6-DoF grasp estimation for articulated objects.
The authors developed a dataset containing valid 6-DoF ground truth grasp poses for articulated objects, as well as photo-realistic kitchen scenes with multiple articulated objects.
CenterArt outperforms the state-of-the-art baseline UMPNet in accuracy and robustness, achieving a 28% higher success rate in 6-DoF grasp estimation, even in complex scenes with noisy depth inputs.
The approach utilizes a center-based object detection method and neural implicit representations to efficiently represent and predict the complete 3D information (6D pose, 3D shape, and joint state) of articulated objects.
Stats
The dataset contains 375,266 grasp labels for 766 object-joint state pairs.
CenterArt is trained on approximately 100,000 RGB-D images of realistic kitchen scenes with multiple articulated objects.
Quotes
"CenterArt is the first approach for simultaneous 3D shape reconstruction and 6-DoF grasp poses estimation of articulated objects."
"CenterArt outperforms the state-of-the-art baseline UMPNet in accuracy and robustness, achieving a 28% higher success rate in 6-DoF grasp estimation, even in complex scenes with noisy depth inputs."