Translucent Object Grasping Using Robot Vision

Grasping translucent objects, such as open containers, poses a significant challenge when using RGB and Depth (RGBD) cameras, primarily due to the presence of cavities in their depth values. The need for effectively grasping translucent containers is especially important in kitchen environments, where easy visibility of the contents inside is essential, particularly for individuals with dementia. This paper addresses this challenge by introducing a novel method that combines an analytical approach with an object detection algorithm such as You Only Look Once (YOLO) to improve grasping performance. Traditional approaches often rely on depth-filling deep neural network models to mitigate the issues caused by these cavities. Although various deep learning methods have been developed for this purpose, they typically entail extensive data collection efforts for fine-tuning their models to work for the objects of interest. In contrast, the approach presented in this paper leverages an analytical method that is particularly well-suited for objects with simple geometries, effectively eliminating the necessity for extensive data collection to predict grasp points and fill cavities. The experimental results demonstrate the effectiveness of this novel approach, with an average grasping accuracy of 94.55% achieved on translucent open containers, establishing it as a viable and practical alternative to traditional deep learning-based methods. The source code is available at Link
^1
1
https://github.com/HMI2-Research-Group/Analytical_BestGrab and the dataset for training the object detection algorithm YOLO in this paper is available at link
^2
2
https://github.com/HMI2-Research-Group/Kitchen-YOLO-Dataset.