SEG-Net: Deep Learning Grasping With a Soft Enveloping Gripper

The emergence of non-fingered soft bioinspired grippers poses a challenge for learning-based grasping control due to the lack of a model describing grasping robustness and a dataset for training. In this letter, we propose a comprehensive pipeline encompassing grasping evaluation, dataset generation, deep neural network construction and training, as well as experimental verification for a soft enveloping gripper to investigate its learning-based grasping methods. The core of our approach lies in the development of a grasping quality model based on force balance between objects and the deformed gripper, enabling us to evaluate the robustness of grasp. Using this model, we synthesize approximately 10K grasp scenes and 30K grasp poses for our dataset, each containing four pixel-wise heatmaps representing depth information, grasp depth, grasp axis, and quality assessment. Subsequently, we construct SEG-Net which takes in depth images as input and outputs the best grasping point along with corresponding grasp axis and depth. Following training and fine-tuning using our dataset, we validate the performance through simulations as well as experiments. Results demonstrate that our proposed method effectively enables automatic grasping using the soft enveloping gripper.