OpenAI’s Learning Dexterity Platform Gives Robotic Hand the Ability to Manipulate Objects Like…

Gripping an object with your hand is an action ingrained into us at a young age — playing with spoons, blocks, and anything else in our immediate area were examined (and eventually taste-tested) when we were toddlers.

By the time we became teenagers we were flipping pens, learning to play musical instruments, and building weight-bearing structures for physics class (while still taste-testing anything around us). The point is, through learning, our hands became more dexterous, which is what OpenAI’s Learning Dexterity platform has done for a robotic hand.

According to the blog, the researchers used their AI learning platform OpenAI Five to train a Shadow Dexterous Robotic Hand to handle objects- pivoting, sliding, gaiting, and more, using a level of anthropomorphic skill similar to humans. The OpenAI Five algorithm employs reinforcement learning for training, which is a process that focuses on attaining a complex objective through a reward/penalization goal system.

The training model (known as domain randomization) utilizes a three-camera setup to follow the robotic fingers in real-time while a computer keeps track of their position. This provides a myriad of different experiences that can be scaled up by the researchers quickly over other training models.

Their system, known as Dactyl, uses a human-like set of approaches to achieve the desired result, such as manipulating a block until the correct face appears at the top. With Dactyl being an AI-based platform, the tasks were not taught to the system but instead learned through the reinforcement learning process.

The Shadow Dexterous Hand (from the Shadow Robot Company) is an excellent piece of technology that pairs well with Dactyl and features 20 actuated degrees of freedom, force sensors that provide accurate position, and ultra-sensitive touch sensors on the fingertips that enable it to grip both hard and soft objects without damaging them. It can also be teleoperated or mounted on robotic arms as part of an overall system.

OpenAI’s Learning Dexterity platform was two years in the making and is still under development, which is incredible for a platform that produced a new learning algorithm, scaling it up to solve complex tasks and applied it to real-world dynamics. Their ultimate goal of the researchers is to continue to scale-up the capabilities of their platform for a “safe AI general intelligence,” and if that’s something you would like to be a part of, they are hiring.