The Experimental Setup

In this section we briefly describe the simulated robot, the architecture of the neural controller, and the adaptive process used to train the robot to grasp objects of different shapes. For details of all parameters used see the paper.

The Robot

The robot used in the experiments reported is a simulated humanoid robot including an anthropomorphic robotic arm with 7 actuated degrees of freedoms (DOFs), a robotic hand with 20 actuated DOFs, proprioceptive and touch sensors distributed within the arm and the hand, and a vision system located in the robot's head.

The robotic hand (Figure 2) is composed of a palm and fourteen phalangeal segments that make up the digits (two for the thumb and three for each of the other four fingers) connected through 15 joints with 20 DOFs.

The Neural Controller

The robot is provided with the neural controllers shown in Figure 2 which include 21 sensory neurons, five internal neurons with recurrent connections and 16 motor neurons.

The object position sensors, arm and hand propriosensors, and tactile sensors encode the state of the corresponding sensors described above. The actuators of the arm encode the activity of the 14 motor neurons controlling the corresponding muscles of the arm. The two actuators of the hand encode the desired extension/flexion state of the thumb and of the other four fingers, respectively (i.e. the four fingers are not controlled independently). See the paper for details.

The Adaptive Process

The free parameters of the neural controller – the connection weights, the biases of internal neurons and hand actuators, and the time constant of leaky-integrator neurons – have been adapted through an evolutionary robotics method with following parameters:

Videos of Robot in Action

Robot 5 grasping different objects