A robot hand solves a Rubik’s cube
OpenAI trained the same Shadow hand to solve a Rubik’s cube, generating progressively harder simulated environments automatically; the success rate was 60% on a half scramble and 20% on a full one.
Why it matters
Where a year earlier a human had to choose how much to randomise, the program now built that curriculum itself, and task length grew from a few seconds of reorientation to a multi-minute sequence. The honest 60% and 20% are also part of why this is a record rather than a reel.
The method is called Automatic Domain Randomization: instead of a fixed randomisation range set by hand, it widens the range as the policy copes. The best policy, ADR XXL, used vision for estimating the cube’s pose and an instrumented Giiker cube for tracking face angles. Over ten trials in each condition it solved a half scramble, requiring 15 face rotations, in 60% of cases, and a full scramble, requiring 26 rotations, in 20%. The manipulation policy was trained only in simulation. The cube-solving algorithm itself is not the point here: the difficulty is in the hand physically executing the required sequence of turns without dropping the cube.