LearningGripper

Gripping and orienting through independent learning

The LearningGripper looks like an abstract version of the human hand. Its actions also have a lot in common with its natural role model. The four-fingered robot gripper uses machine learning to independently learn to rotate a ball in a raised position in any given orientation.

Only the elementary actions and possible positions of the LearningGripper’s fingers, plus the function for feedback from the environment, are defined in advance so that this complex action can be carried out. The gripper is only told what it needs to be able to do, not how to do it. The movement strategy required for this is developed independently by the gripper’s learning algorithms – without any further programming.

Gripping and learning – intelligent interaction

Theories state that we humans are only so intelligent because our hands can solve so many complex tasks. Babies start to grasp objects very early – for example, their mother’s finger. Once we have learned to grasp an object correctly, we can turn it around and look at it from all sides. This is the only way a 3D image of the object can be reconstructed in the mind. So the hand also helps us humans to learn.

Trial and error – learning through reinforcement

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. The LearningGripper uses the method of reinforcement learning. The gripper is not given any concrete action that it has to imitate. It optimises its skills solely on the basis of feedback on its previous actions. This increases the probability that it will execute a successful action and will not repeat a less successful move.

The four fingers are driven by a total of 12 pneumatic bellows actuators with low pressure between 2.5 and 3.5 bar. Each of them has three degrees of freedom and the basic functions of the index finger. In the initial state alone, the entire hand thus has a choice of 3¹² total actions to reposition the ball. Due to the intelligent coordination of the fingers and the flexible bellows structure, the kinematics are free-moving and pliable. It can safely grip, lift and rotate even the most sensitive objects – just like its natural role model.

Reduced programming effort thanks to machine learning

At the LearningGripper trade fair exhibit, a gripper demonstrates how it takes less than an hour for it to learn a mechanical motion strategy – from its first attempt to the reliable execution of the required task. A second gripper shows the learned procedure in the desired target scenario. It lifts the ball and turns it so that the embossed lettering on the end is visible in the centre on the top.