OpenAI’s Dactyl system uses OpenAI Five neural networks to train a Shadow dexterous hand how to solve a Rubik’s Cube. (Image credit: OpenAI)
More than a year ago, San Francisco-based AI research lab OpenAI managed to train a robotic hand to manipulate a cube with human-like ability, and now the company has reached another milestone in its endeavor to build self-learning robots by using that robotic hand to solve a Rubik’s Cube. OpenAI explains, “Since May 2017, we’ve been trying to train a human-like robotic hand to solve the Rubik’s Cube. We set this goal because we believe that successfully training such a robotic hand to do complex manipulation tasks lays the foundation for general-purpose robots.”
OpenAI’s Dactyl system uses the company’s OpenAI Five reinforcement learning algorithm, along with a new technique known as ADR (Automatic Domain Randomization), which generates progressively more challenging environments in simulation. Dactyl trains the Shadow dexterous robotic hand how to manipulate the cube using neural network-based simulations, which then transfer over to real-world actions.
To train the robotic hand through those progressively harder simulations, OpenAI needed some serious hardware, so they built a processing platform that uses 6,144 CPUs and 8 Nvidia V100 GPUs. Using the hardware allows Dactyl to pack 100 years’ worth of training simulations into 50 hours. As for the Shadow robotic hand, it was modeled after humans and offers 24 joints, has 20 actuated degrees of freedom, position and force sensors, and ultra-sensitive touch sensors on the fingertips.
OpenAI states that using their current simulation methods, the robotic hand can only manage to solve the Rubik’s Cube 20% of the time when using a maximally challenging scramble that requires 26 face rotations. As you might imagine, lowering the difficulty results in higher success rates, and while that doesn’t sound too impressive, it’s a significant step in creating autonomous robots capable of learning problem-solving skills on their own.
Have a story tip? Message me at: cabe(at)element14(dot)com