In just one year, the AI research startup went from only getting the robot to manipulate the entire cube in-hand to manipulating the position of the colors and solving the Rubik's Cube 60% of the time. Using simulation, the researchers tasked the robot with solving the cube in progressively harder conditions, such as increasing the size of the cube. The team successfully tested the physical hand with multiple perturbations, including tying fingers together and placing a rubber glove over the hand – all scenarios that the neural network model was never trained on in simulation. While most applications of reinforcement learning have only applied to structured environments like games, this robotic application could be applied in industry.
For the original news article,