Generally, there are three criteria that define how well a humanoid robot hand resembles human hand: shape (anthropomorphism), dexterity (degree of freedom) and sensation (sensor). In fact, it is not easy to fulfill three requirements all at the same time. An anthropomorphic robot hand may not have powerful grasping force due to space, geometry constraints. A highly sensitive robot hand requires tactile sensor installed in each finger link and signal processing device to process data. However, this robot hand is prone to hazardous environment, moisture, and vibration.
Balancing between three criteria, we designed humanoid robot hand that is anthropomorphism, optimized for grasping task and light weight. We emphasized grasping task over other tasks because it is main and common function of human hand. We invented a novel mechanism which combines pulley, belt and torsion spring to create under-actuated behavior. Rather than using tactile sensors or force torque sensors, we devised an observer algorithm to predict grasping force. This method eliminates the need of A/D converters, current sensors, analog amplifiers, filters and signal cables, which are the source of maintenance, calibration problem and increases in weight and size of robot hand.
To verify our under-actuated mechanism and observer algorithm, a two-finger robot hand is controlled by a host PC through serial communication. The experiment result is quite promising. The robot hand can hold light object in different geometries like sphere, cylinder, square box, cone, thin plate, etc. Finger can sense the external force thus it knows when it contacts to an object and maintains grasping force as desired.