"GraspVar-5: A Multi-Variant Dataset for Learning-Based Robotic Grasping"

"Robotic grasping in real-world environments requires robust perception models capable of handling variations in object geometry, size, and material properties. Existing datasets primarily emphasize large-scale data collection, often overlooking structured intra-class variability critical for generalization. This paper introduces GraspVar-5, a compact yet systematically designed dataset for vision-based robotic grasp detection.The dataset consists of five everyday object categories\u2014plastic bottle, puzzle cube, tennis ball, plastic ball, and sponge ball\u2014with 200 variations per category, capturing controlled differences in shape, size, and material characteristics. The dataset includes both contact point annotations and bounding box labels, enabling supervised learning of grasp configurations. Images are captured using a camera-based setup and further augmented to enhance diversity. For efficient training, images are converted into NumPy array format.GraspVar-5 is specifically designed to support learning-based grasp models such as GGCNN and provides a benchmark for evaluating grasp robustness under controlled variations. The dataset is publicly available and aims to facilitate research in robotic manipulation, assistive systems, and industrial automation. "