Title:DexWrist: A Robotic Wrist for Constrained and Dynamic Manipulation

Abstract:Development of dexterous manipulation hardware has primarily focused on hands and grippers. However, robotic wrists are equally critical, often playing a greater role than the end effector itself. Many conventional wrist designs fall short in human environments because they are too large or rely on rigid, high-reduction actuators that cannot support dynamic, contact-rich tasks. Some designs address these issues using backdrivable quasi-direct drive (QDD) actuators and compact form factors. However, they are often difficult to model and control due to coupled kinematics or high mechanical inertia. We present DexWrist, a robotic wrist that is designed to advance robotic manipulation in highly constrained environments, enable dynamic and contact-rich tasks, and simplify policy learning. DexWrist provides low-impedance actuation, low inertia, integrated proprioception, high speed, and a large workspace. Together, these capabilities support robust learning-based manipulation. DexWrist accelerates policy learning by: (i) enabling faster teleoperation for scalable data collection, (ii) simplifying the learned function through shorter trajectories and decoupled degrees of freedom (DOFs), (iii) providing natural backdrivability for safe contact without complex compliant controllers, and (iv) expanding the manipulation workspace in cluttered scenes. In our experiments, DexWrist improved policy success rates by 50-55% and reduced task completion times by a factor of 3-5. More details about the wrist can be found at this https URL.