Announcing major progress towards general-purpose robots. This robot is trained using the same reinforcement learning code as OpenAI Five, plus a new technique for transferring knowledge from simulation to reality. Result is unprecedented dexterity:
OpenAI
 @OpenAI
15 Oct 2019
We've trained an AI system to solve the Rubik's Cube with a human-like robot hand. This is an unprecedented level of dexterity for a robot, and is hard even for humans to do. The system trains in an imperfect simulation and quickly adapts to reality: openai.com/blog/solving-rubi…

Oct 15, 2019 · 4:01 PM UTC

9
101
10
524
Replying to @gdb
Up next: one handed brain surgery! Great job OpenAI, bringing RL to handle this type of problem has been a long time coming. I think all the hand-made (haha) approaches are looking less and less attractive compared to RL. This has great immediate future in prosthetics too.
Replying to @gdb
I love this
Replying to @gdb
Amazing step after learning dexterity. And the paper seems to be really amazing with 51 pages!
2
Replying to @gdb
This is general purpose in what sense? I still see that it specializes in solving Rubik’s cube no?
1
Replying to @gdb
What type of memory-augmented neural network is used? From the memory probing section, it looks like slots of some type of neural Turing machine ?
Replying to @gdb
Great progress in robot mechanics: fine grained dexterity and efficient sim2real will be crucial for embodied AI. Once key intelligence mechanism are discovered, solutions like this one will convert plans into actions and adjust accordingly