Our robotic system solves the cube 60% of the time under normal conditions — but only 20% with an adversarially-scrambled cube. The big result is that it's possible at all. Like with OpenAI Five, reliability keeps getting better the more we train
This tweet is unavailable

Oct 15, 2019 · 6:51 PM UTC

5
4
2
48
Replying to @gdb
Any task is possible with Deep Learning given enough training time and compute power, but it won’t be transferable to other domains. The agent still has no knowledge of reality like living things with brains.
1
2
Replying to @gdb
Are you running any other systems to solve diverse problems like this in parallel rn or in the future?
Replying to @gdb
Are you defining an official Rubik's cube scramble as "adversarial"? Should you not coin your half scramble as "easy" instead?
5