Although human hands have lots of degrees of freedom, we typically don’t use most configurations. For example, we usually don’t move the last two joints of our fingers independently. Now let’s look at the anthropomorphic robot hand below. Like the human hand, it has lots of degrees of freedom and planning a motion would typically take a lot of time if we consider all possibilities. To solve this problem, Rosell et al. propose to look at what motions humans do, and use the information to limit the motions the robot hand should be doing.
To learn about human hand motion they fitted a human with a sensorized glove and recorded its movements. The human movements were then translated into robot coordinates. Using a technique called Principal Component Analysis, the robot is able to extract the most important motions that humans do. By combining these principal motions with a planner to make sure the arm and hand don’t collide with the environment or their own parts, the robot is able to perform human-like motion using little computation.
The approach was validated in simulation and using a four finger anthropomorphic mechanical hand (17 joints with 13 in- dependent degrees of freedom) assembled on an industrial robot (6 independent degrees of freedom).