Reinforcement Learning With a Low‐Cost Parallel Linkage Gripper for Static and Dynamic Planar In‐Hand Manipulation

Parallel linkage grippers provide a low‐cost and compact alternative to anthropomorphic hands for planar manipulation tasks. However, their closed‐chain kinematics introduce actuator coupling, singularity‐related constraints, and position‐dependent manipulation capability, making closed‐loop control challenging. This work presents D‐PALI‐E (Dual PArallel LInkage Enhanced), an open‐source, predominantly 3D‐printed parallel linkage gripper developed through hardware‐learning co‐design and equipped with passive‐joint limiters inspired by the overextension‐limiting function of the human finger volar plate. During contact, the limiters serve as motion constraints and part of the load‐bearing structure, modifying the gripper’s mechanical operating characteristics. They exclude mechanically undesirable configurations, participate in load transmission during pinch grasping, and reshape the operating region for policy learning. Reinforcement learning is used for quasi‐static pose adjustment (reorientation, repositioning, and full pose rearrangement) and dynamic pinch grasping (catching rolling balls of various sizes and directions and capturing slow or stationary objects). Sim‐to‐real transfer is achieved using system identification and domain randomization. Experiments demonstrate pose adjustment for regular geometric objects (within 5 mm and 5°), and a 72.5% average success rate in pinch grasping of rolling balls. These results show that the combined mechanical redesign and learning framework support both quasi‐static planar pose adjustment and dynamic pinch grasping on low‐cost hardware.