?utm_source=dlvr.it&utm_medium=twitter

Humans perform visual recognition/detection and then predict motion of independent objects. The brain relies strongly on understanding of physical world and object geometry/affordances in estimating motion (and thus it's much more difficult), the "one model fits all" approach is completely wrong. Shouldn't the baseline be detection - motion estimation? If you are really interesting in solving the robotic arm problem why not just create a network that leverages ground truth information about location and motion of the arm.