Simple tasks, like choosing the right object and hitting it, are still too difficult for robots to perform, as the solution requires complex sensorimotor interactions between the robot and its environment. But thesolution is the same as for the humans: learning by doing is the key to it. Ali Ghadirzadeh et al. from KTH Sweden developed a data-efficient deep predictive policy training (DPPT) framework with a deep neural network policy architecture. An experiment with a PR2 robot demonstrates the effectiveness of this framework by needing only 180 attempts to learn skilled object grasping and ball throwing. Find the article here.