Image-based grasping for soft-hands

This paper addresses the challenge of autonomous grasping and manipulation of delicate archaeological artifacts using a soft robotic hand, designed to minimize damage while maximizing adaptive handling capabilities. Leveraging the Pisa/IIT SoftHand, this research integrates deep learning with robotics to predict optimal grasp poses for fragmentary historical objects. Utilizing Isaac Gym for simulation, an extensive dataset was generated, capturing multi-view images and corresponding successful and unsuccessful grasp poses. This data supports a two-stage neural network: a pose estimator followed by a grasp success predictor network, that complement each other to just focus on using the poses with the highest probability. Techniques such as data augmentation, dataset balancing, and transfer learning on limited samples aim to improve model generalization in real-world scenarios. This research contributes to the RePAIR Project’s broader goal of artifact restoration and preservation, illustrating how machine learning and robotics can facilitate artifact handling in archaeology, thus advancing heritage conservation technologies.