MIRROR
IST–2000-28159
Mirror Neurons based Object Recognition
|
MIRROR IST–2000-28159 Mirror Neurons based Object Recognition |
DIST -- University Of Genoa
Learning gravity compensation.
- Video1: arm zero-weight. The robot keeps the arm in a stationary position; the closed loop control is not activated and only the arm’s weight is compensated by the controller. The arm can be moved by a human as if it was very light.
- Video2: low stiffness. In this case the arm is controlled and randomly moved to show the low-stiffness control. A human can safely interact with the manipulator.
Learning the hand model:
- Video 1,2: hand segmentation. The robot moves the arm around and performs a periodic motion of the hand. Motion in the image plane is correlated with the periodic motor command in order to get a segmented image of the hand. The video shows that the segmentation is not influenced by external movements.
- Video 3: hand localization. Right window: the color histogram of the hand is used to segment the hand from the background. Brighter pixels have higher probability to belong to the hand. The color histogram is not able to discriminate between objects with a similar color (the yellow car). Left window: a neural network is trained to predict the position of the hand from the arm posture (the red cross superimposed to the image). In this case no-visual information is actually used. By putting together this information with the histogram probability, the robot is able to localize the hand more efficiently.
Hand:
- Video 1: hand compliance. The video show the intrinsic elasticity in the hand mechanics. The hand here is not controlled.
- Video 2,3,4: grasping different objects. To test the hand mechanics some objects are grasped. In this case the hand is remotely controlled by an operator. It is important to note how the fingers adapt to the shape of the object being grasped.
Grasping strategies:
- Video 1: a simple grasp reflex-like mechanism is implemented. Whenever pressure is applied to the palm of the hand, a grasping movement is initiated. After a certain period of time the hand is opened again and the object released.
IST - Istituto Superior Tecnico in Lisbon
Model of mirror neurons:
- Visual processing related to the modeling of mirror neurons is shown here. The first video shows an example of the precision grip. Images here were recorded using the data-glove setup developed especially for Mirror. This is exactly the type of visual information that was processed by the Bayesian classifier described in deliverable D3.4. Video number 2 shows an example of power grasp under the same experimental conditions. Video number 3 shows how the hand is located (based on color) and its appearance mapped to a standard reference frame. The normalized and rescaled pictorial information is successively processed by the PCA algorithm.
DP - University Of Uppsala
Rotating rod experiment
Infants' ability to adjust hand orientation when grasping a rotating rod has been studied. The rod to be reached for was either stationary or rotated. The results show that reaching movements are adjusted to the rotating rod in a prospective way and that the rotating rod affects the grasping but not the approach of the rod.
video: (1)
DBS - University Of Ferrara
In-vivo recordings of mirror neurons in behaving monkeys
Different classes of neurons are recorded during grasping action in different conditions of "visual feedback". The videos show grasping actions performed in four different conditions: 1) with ambient illumination; 2) in the dark; 3) with a flash of light at the instant of maximum finger aperture; 4) with a flash of light at the instant of touch.