Festo_BionicSoftHand_en

02: Machine vision: computer vision to collect the necessary data for a vir- tual image 05: Massively parallel learning: fast learn- ing through the duplication of the digi- tal twin BionicSoftHand Pneumatic gripper based on the human hand Whether grasping, holding or turning, touching, typing or press- ing – in everyday life, we use our hands as a matter of course for the most diverse tasks. The human hand – with its unique combination of force, dexterity and fine motoric skills – is a true miracle tool of nature. An important role is played by the human thumb, which is positioned opposite the other fingers. This so- called opposability enables us, for example, to clench a fist, to grasp precisely and to also do filigree work. Compliant kinematics for safe collaboration What could be more logical than equipping robots in collaborative working spaces with a gripper that is modelled on this natural model and can learn through artificial intelligence to solve various tasks? The BionicSoftHand is pneumatically operated so that it can inter- act safely and directly with people. Its gripper fingers consist of flexible bellows structures with air chambers and other soft mater- ials. This makes it light, flexible, adaptable and sensitive, yet cap- able of exerting strong forces. Functional integration in the tightest of spaces In order to carry out the movements of the human hand realistic- ally, small valve technology, sensor technology, electronics and mechanical components are integrated in the tightest of spaces. Gripping and learning – intelligent interaction By means of artificial intelligence, the bionic robot hand learns to independently solve gripping and turning tasks similarly to the human hand in interaction with the brain: our hands not only react to the commands of the brain but also simultaneously provide it with important information to adapt further actions to the environ- ment and its requirements. Neuroscientists say that humans are only so intelligent because the hand can solve so many complex tasks. Babies start to grasp very early – for example, the mother’s finger. Once they have learn- ed to grasp an object correctly, they can rotate it and look at it from all sides. This is the only way a 3D image of the object can be reconstructed in the head. Thus, the hand also helps humans to learn. Reinforcement learning: the principle of reward The learning methods of machines are comparable to those of humans – be it positive or negative, they both need feedback on their actions in order to classify them and learn from them. BionicSoftHand uses the method of reinforcement learning, learning by strengthening. This means that instead of having to imitate a concrete action, the pneumatic robot hand is merely given a goal. It tries to achieve this through trial and error. Based on the feedback received, the hand gradually optimises its actions until it finally solves the task successfully. Digital twin of the real robot hand Specifically, the BionicSoftHand should rotate a 12-sided cube so that a previously defined side points upwards at the end. The necessary movement strategy is taught in a virtual environment with the aid of a digital twin, which is created with the help of data from a depth-sensing camera via computer vision and the algo- rithms of artificial intelligence. Fast knowledge transfer through massively parallel learning The digital simulation model accelerates the training considerably, especially if you multiply it. In so-called massively parallel learn- ing, the acquired knowledge is shared with all virtual hands, which then continue to work with the new state of knowledge: so each mistake is made only once. Successful actions are immediately available to all models. After the control has been trained in the simulation, it is trans- ferred to the real BionicSoftHand. With the virtually learned move- ment strategy, it can turn the cube to the desired side and orient other objects accordingly in the future. Learning algorithms instead of complex programming In automation today, many tasks are too complex to be able to dir- ectly program every movement and function. Due to its many de- grees of freedom, conventional control strategies are not readily applicable with the BionicSoftHand. In order to fully exploit its productivity and efficiency potential, it needs to learn on its own how to adapt its behaviour and, subsequently, expand its skills. 01: Complete pneumatic system: safe interaction with the BionicSoftHand on the BionicSoftArm 03/04: Digital twin: the real robot hand and its virtual image in the simu- lation model 01 02 05 04 03 2 Festo AG & Co. KG 3 BionicSoftHand: Pneumatic robot hand with artificial intelligence