Robotic grasping has undergone transformative progress, moving from fixed‐motion stiff grippers to adaptive systems capable of cognitive interaction. Among the myriad of systemic options, soft robotic grippers, with their inherent compliance and adaptability, offer a promising route for overcoming the challenges of manipulating deformable, irregular, or fragile objects. Augmenting these systems with learning‐based control and perception strategies could lead to robust, generalizable, and intelligent manipulation. While progress in both soft robotics and machine learning has been significant, there is a gap in the literature relating to the intersection of these two domains. This review bridges that gap by providing a pipeline‐oriented perspective on learning‐based soft robotic grasping. We examine recent advances in soft gripper design, multimodal sensing, learning‐based planning, and control strategies. We also provide a summary of neural network architectures used in grasping, explore benchmark datasets, and highlight generalization problems in unstructured environments. We go on to outline key challenges and potential development pathways, emphasizing the need for self‐supervised learning, sim‐to‐real transfer, and unified architectures that combine physical modeling with adaptive data‐driven policies. This survey is intended to be a valuable tool and backdrop for developing scalable, robust, and lifelong learning‐enabled soft robotic grasping systems.
Majumder et al. (2026) studied this question.