OpenAI Builds a Robot That Can Single-Handedly Solve a Rubik’s Cube – Literally

If you’re anything like us, you think a Rubik’s cube is an intellectual torture device that can only be solved with the use of dark magic or Faustian bargains made at crossroads. And yet, many people are capable of solving the puzzle so quickly that it looks almost easy. Even robots can solve them through the methodical application of known techniques. But most of those robots are design specifically for that task. This robot built by OpenAI is able to solve a Rubik’s cube using just one hand that is normally intended for general purpose use.

OpenAI isn’t showcasing this robot to prove that they can train a robot to solve Rubik’s cube — that’s old news. The purpose of this project is to demonstrate how dexterous a robotic hand can be, and how a robot can use that with confidence and accuracy. The robotic hand itself is a Shadow Dexterous E Series Hand, which OpenAI used in another demonstration that we featured last year. Solving the Rubik’s cube is an extension of that previous work, which utilizes virtual simulations to provide an environment for the robot to quickly learn how it can interact with physical objects in the real world.

In order for any artificial intelligence system to learn how to perform a task, it needs training. That can be done in the real world, but it’s a slow process. Virtual simulations can be run many times faster than real-time, and can even be run concurrently. OpenAI’s “sim2real” system lets the robot experiment and learn through trial and error in a simulated environment. In order to accommodate the physics and unpredictable nature of reality, sim2real pushes a lot of randomness into the simulations. That gives the robot enough experience with imperfect scenarios to be able to solve the Rubik’s cube reliably. As you watch, it seems able to manipulate the Rubik’s cube without any hesitation at all. It shudders a bit because of the motor movement, but otherwise has more finesse than most humans do.