Putting on glasses, raising a coffee cup to the mouth or a typing an SMS – for the human hand it is easy to grip and move objects. Even newborn children begin to hold objects or the fingers of their parents. Gradually and by trial and error they learn how to hold, move and turn an object with their hands so they can look at it from all sides and form a three-dimensional impression of it.
Gripper learns by trial and error
The LearningGripper from Festo behaves in just the same way. This is a gripper with four fingers, inspired by a human hand. With the help of machine learning software, this gripper can learn a complex action like picking up and orienting an article – in the example shown here, a ball with a printed logo. All that needs to be defined in advance is the basic positions of the fingers and the feedback function from the environment; the gripper learns all other motion sequences by trial and error.
Points as a reward
Its task is to turn the ball until the logo is at the top. At the beginning the LearningGripper tries to move the ball randomly. A position sensor in the ball provides feedback on how far the logo is from the gripper's "palm". The greater the distance, the stronger the positive feedback, and the LearningGripper is given a reward based on a points system. This is processed in the machine learning software. Over a period of time, the software develops a movement strategy and the gripper learns what action to take at a particular point. It changes its motions in such a way that it gets as much positive feedback as possible and finally finds a reliable solution to its task. If the strategy of one gripper is transferred to another, it uses that as a knowledge base which will enable it to learn its own strategy more efficiently.
Potential for the factory of the future
With its LearningGripper, Festo is showing how systems in the future will be able to solve complex tasks autonomously without complex programming. Self-learning systems like the LearningGripper could be installed on a production line in the future and would then optimise their behaviour independently.
A film on the LearningGripper shows how it carries out tasks: