If a robot arm is to grasp an object, it often fails in practice when it is not able to access the target object directly. It can take it minutes to calculate which indirect path to take – by first moving an obstacle out of the way, for example. Researchers at the University of Leeds, England, have now combined two AI techniques to remedy the situation: automated planning and reinforcement learning. The first simulates the possible sequence of moves the gripper arm could execute to reach the target object. The latter trains it through a sequence of thousands of trial and error attempts to reach and move objects through which it learns which actions are more likely to end in success. According to the university, the robots consequently develop an ability to generalize, which should speed up their ‘thinking’ time by a factor of ten.

Reinforcement learning in particular is currently the subject of intensive research. In an interview with German electrical engineering and automation portal elektrotechnik AUTOMATISIERUNG , Manuel Kaspar from robotics provider KUKA’s research team describes it as a “promising version of machine learning”. The catch for industry, however, is that it is almost impossible in industrial processes to let the system fail as often as desired until it finally grasps a situation correctly.