Link to paper ( why no Arxiv :/ ): https://d4mucfpksywv.cloudfront.net/research-covers/learning-dexterity/learning-dexterity-paper.pdf https://d4mucfpksywv.clou

Link to paper ( why no Arxiv :/ ): https://d4mucfpksywv.cloudfront.net/research-covers/learning-dexterity/learning-dexterity-paper.pdf https://d4mucfpksywv.cloudfront.net/research-covers/learning... TLDR (quick-ish skim, feel free to correct) they train a deep neural network to control a robot hand to choose desired joints state changes (binned into 11 discrete values; eg rotate this joint by 10 degrees) for a 20-joint hand given low-level (non-visual; so, current and desired 3D orientation of the object and exact numeric state of the joints) input of the state of a particular object and the hand. They also train a network to extract the 3D pose of a given object given RGB input. All this training is done in simulation with a ton of computation, and they use a technique called domain randomization (changing colors and textures and so friction coefficient and so on) to make these learned models pretty much work in the real world despite being trained only in simulation. It's pretty cool work, but if I may pull my reviewer hat on not that interesting in terms of new ideas - still, it's cool OpenAI is continuing to demonstrate what can be achieved today with established RL techniques and nice distributed compute.