A Quasi-Direct-Drive Underactuated Asymmetric Hand for Dexterous and Efficient Grasping and Manipulation
Manuscript / preprint
Abstract
We present the Berkeley QUAD (Quasi-direct-drive, Underactuated, Asymmetric Design) Hand, a four-finger anthropomorphic robotic hand with 11 degrees of freedom and 8 degrees of actuation. Quasi-direct-drive actuation at the base of each finger enables high force transparency for dexterous, adaptive performance. We address the size, weight, and thermal limits of these actuators through bio-inspired asymmetry: dexterity is delegated to the independently actuated radial fingers, while a larger actuator and compliant linkage provide strength and passive adaptation on the ulnar side. Hardware evaluations demonstrate the efficiency of this reduced-DoA, asymmetric design: it achieves 29 of 33 Feix taxonomy grasps, backdrive forces as low as 50 g, and up to a 96-fold reduction in heat generation during sustained loading.
Design
Most robotic hands actuate each finger indepentently and identically (or nearly so). This leads to a common compromise across weight, strength, and sensitivity, with different designs favoring one over the others. Our design challenges this assumption -- do all fingers need the same capabilities, or can we achieve better performance by specializing them? We take inspiration from human hand function, providing the radial fingers with independent QDD actuation for fine dexterity. On the other (side of the) hand, the middle and ring fingers share a larger, more efficient motor through a compliant transmission. This saves palm volume and, after contact, passively sweeps the ring finger inward for stronger grasp closure.
Interactive grasps
Explore how the ulnar fingers adapt to different object geometries.
Drag to orbit and use the playback controls to scrub through each grasp.
Results
Hardware experiments demonstrate more practical outcomes of our design. With QDD actuation, all fingers are force-sensitive, with the radial fingers reaching backdrive forces as low as 50 g. This sensitivity enables delicate manipulation and precise control at the cost of high power consumption and heat generation during sustained loading. The ulnar fingers are less sensitive but more powerful and efficient, supporting firm grasps and large static loads over time. Specifically, we find that the shared actuator design reduces expected heat generation by up to 96 times compared to the independently actuated radial fingers—even when considering a load double in size. In addition to efficiency, the hand broadly retains dexterity, achieving 29 of 33 Feix grasps and 10 of 11 Kapandji opposition postures. This is enabled by our novel ulnar mechanism, which passively preserves human-like grasp closure with fewer degrees of actuation.
Grasp diversity
Citation
Acknowledgments
This work was supported by the Berkeley AI Research Center for Humanoid Intelligence and the National Science Foundation Graduate Research Fellowship Program under Grant No. 2146752. We thank members of the Embodied Dexterity Group, Nathan Davis, Alex Liang, and General Motors for their support and contributions.