Not to be too negative, it's cool work, but I'd argue unlike the OpenAI result it is not so surprising this was doable with the techniques they used ; see eg th

Not to be too negative, it's cool work, but I'd argue unlike the OpenAI result it is not so surprising this was doable with the techniques they used ; see eg this paper from Google http://www.roboticsproceedings.org/rss14/p10.pdf http://www.roboticsproceedings.org/rss14/p10.pdf and this one from Stanford/DeepMind http://www.roboticsproceedings.org/rss14/p09.pdf http://www.roboticsproceedings.org/rss14/p09.pdf . Yes there is the additional aspects of an object in hand, but fundamentally the techniques are the same. Of course these works are cited in related works of paper as they should be; perhaps the OpenAI blog should also provide more context on where this stands wrt prior work, as many non-researchers may read this is may be quite misleading...