The OpenAI has become a new milestone making robots capable of solving a Rubik’s Cube one-handed. The OpenAI-based company has trained neural networks in simulation using reinforcement and learning to make this achievement possible. The company has been working on the OpenAI project since May 2017. Recently the company has achieved its goal and marking this as its milestone towards its progress in their field.
OpenAI’s AI-powered robot
The time taken by the OpenAI robotic hand varies depends on how the cube is to be shuffled. On average, it takes nearly about three to four minutes to solve the cube. However, note that this is not the first-ever robot that managed to solve the Rubik’s cube. A robot developed by the MIT graduate students solves the cube in 0.38 seconds. However, that robot doesn’t solve the cube single-handed. Instead, it uses individual motors to actuate each face of the cube.
What sets OpenAI’s robot apart from the robot developed by MIT is that OpenAI’s robot is designed to be a multipurpose robot. On the other hand, MIT robot is explicitly designed to solve that one particular use-case rendering it unusable for performing different tasks.
OpenAI uses a technique called Automatic Domain Randomization (ADR) while training this new robotic hand. “ADR starts with a single, nonrandomized environment, wherein a neural network learns to solve Rubik’s Cube. As the neural network gets better and better at the task and reaches a performance threshold, the amount of domain randomization is increased automatically.”, wrote OpenAI in a blog post.
The company has also mentioned that their neural network is more likely to fail in their first few flips and rotation. The neural network is because the neural network has to find the balance between solving the cube and “adapting to the use and the physical world” in those early rounds.
Take a look at the video of OpenAI’s robot hand solving the Rubik’s cube below and drop your comments in the comment section below about your thoughts on it.