Researchers of the AI initiative OpenAI have one Robotic arm developedwho can handle various fine motor tasks. A first success: The machine has a Rubik's Cube – also called magic cube – solved with one hand and independently. The basis is a machine learning program based on the principle of Reinforcement Learning. This controls fingers and joints of the hand.
In order to achieve this, the researchers faced a challenge: how can software virtually practice parallel solution approaches? For this purpose, a virtual environment was programmed that simulates physical laws of physics. There, a model of the robot hand and its sensors is used. Such environments can then run in parallel thousands of times. The program learns so much faster, how to deal with the magic cube.
The training environments are created by the Automatic Domain Randomization (ADR) algorithm. This generates continuously heavier and random scenarios from step to step, which the system has to master. A random parameter is about the size of the magic cube.
Success is not guaranteed
The researchers then transferred this model to the real world. There, they also ran the robotic arm under different conditions. In an attempt he has two fingers tied together, sometimes he has a rubber glove on, sometimes the magic cube is touched by external influences. The goal is to make the arm as resistant to external impact as possible. In a future practical use in the real world, that would be essential.
The Rubik's Cube still does not solve the arm completely reliably. Twenty percent of all time-limited attempts lead to success when the colors of the cube are mixed for the worst case and 26 turns are necessary. In an average scenario, the robot manages 60 percent of all attempts with success.