02: Visual object detection: a depth camera on the wrist Further development with an optimized radius of action In order to extend the range of the thumb and index finger, the developers have significantly increased the lateral swivel space of both fingers. This means that they can now work together optimally and grip very precisely. Thanks to a 3D-printed wrist with two degrees of freedom, the hand can now also move back and forth as well as to the left and right. This means that gripping with a narrow radius is also possible. Finely tuned gripper with fingertip sensitivity For more stability in the fingers, the air chambers now contain two structural elements that act as bones. A bending sensor with two segments per finger determines the positions of the fingertips. In addition, the hand wears a glove with tactile force sensors on the fingertips, the palm and the outside of the robot hand. This allows it to feel the texture of the object to be gripped and adapt its gripping force to the object in question – just like we humans do. In addition, a depth camera is located on the inside of the wrist for visual object detection. Object detection by means of a neural network With the help of the camera images, the robot hand can recognise and grip various objects, even if they are partially covered. After appropriate training, the hand can also assess the objects on the basis of the recorded data and thus distinguish good from bad, for example. The information is processed by a neural network that has been trained in advance with the help of data augmentation. Extensive data sets through data augmentation In order to achieve the best possible results, the neural network needs a lot of information with which it can orient itself. This means the more training images are available to it, the more reliable it becomes. Since this is usually time-consuming, automatic augmentation of the database is a good idea. This procedure is called data augmentation. By marginally modifying a few source images – with different backgrounds, lighting conditions or viewing angles, for example – and duplicating them, the system obtains a comprehensive data set with which it can work independently. 01: Reliable assistance system: ballbot, DynaArm and BionicSoftHand 2.0 in action 03: Digital learning: data augmentation for training the neural network 03 02 3 BionicMobileAssistant: Mobile robot system with pneumatic gripping hand
RkJQdWJsaXNoZXIy NzczNDE0