OpenAI’s Robot Learns to Solve a Rubik’s Cube With One Hand: Peter Welinder, Research Lead, OpenAI, Shares Insights


In the latest news, OpenAI has developed a robotic arm that can solve the Rubik’s Cube with one hand. This robot was developed by the company with the motive to achieve its goal to build general-purpose robots that can do a wide variety of tasks. Peter Welinder, research lead, OpenAI, shares more about the team’s journey.

OpenAI, an artificial intelligence research organization that develops self-learning robots, recently succeeded in the development of a robotic arm, called Dactyl, that can solve the Rubik’s Cube puzzle. This success brings the company a step closer to its goal of developing a robot that can learn to perform real-world tasks, without the need for months of training.


In this exclusive conversation with IT Toolbox, Peter Welinder, research lead, OpenAI, shares more details about the robot, right from its inception to the final goal of its development.

Peter Welinder is a research lead at OpenAI, where he leads projects on learning-based robotics. Previously, he led machine learning teams at Dropbox and led Anchovi Labs as the CEO and co-founder. Peter has a Ph.D. in Computation and Neural Systems from Caltech and a degree in Physics from Imperial College London.

Here are the edited excerpts from the exclusive conversation with Peter Welinder
on Dactyl’s abilities:

First of all, congrats on the success your team had with Dactyl. Tell us more about the robot. How is the model of the robot hand designed, and which AI techniques does it use to solve the Rubik’s Cube?

Peter: Thank you! Our robot system, Dactyl, is made up of a physical robot hand and two deep neural networks. The first network perceives the configuration of the Rubik’s Cube. We then use a planner, called Kociemba’s algorithm to decide the order in which to turn the faces of the cube. The second network takes the output of these two components, the first network, and the planner, to decide how to move the hand to manipulate the cube. The network is trained from scratch in a simulator, having learned the moves to manipulate the cube entirely by itself.

Could you tell us more about how the idea was conceptualized? How was the journey?

Peter: Human hands let us solve a wide variety of tasks. For the past 60 years of robotics, hard tasks that humans accomplish with their fixed pair of hands have required designing a custom robot for each task. We wanted to see how far we could push deep learning techniques for robotics and picked solving Rubik’s Cube as an example of a very hard manipulation task that hadn’t been accomplished before.

Learn More: What are the Types of Artificial Intelligence: Narrow, General, and Super Opens a new window

What was the motive behind developing this robot, and what is OpenAI’s ultimate goal?

Peter: At OpenAI, we want to build safe general artificial intelligence. We use robots as a testbed for developing general intelligence algorithms. Eventually, we’d like to build general-purpose robots that can do a wide variety of tasks.

What kind of approach did you use to train Dactyl’s algorithm? Were there any obstacles in the way? If yes, how did you overcome them?

Peter: We train our networks in a simulator. The tricky thing with simulators is that they’re never a perfect copy of the real world. If you’re not careful, the algorithm ends up learning strategies that only work in the simulator but not in the real world. For this reason, there’s a lot of manual work that goes into setting up the simulator properly. To overcome this, we developed a technique called Automatic Domain Randomization (ADR). It automatically generates billions of different simulated environments, each one a little different from the others. This forces the network to learn robust strategies that also end up working on a real robot. We’ve found our system to be very robust to changes in its environment. For example, we can tie two of the fingers, or put a rubber glove on the robot hand, and it can still solve Rubik’s Cube.

Learn More: Top 5 Artificial Intelligence Books to Read in 2019Opens a new window

What does the future hold for general purpose, self-learning robots?

Peter: We believe that general-purpose robots could be immensely useful for the world, for example, as assistants to the elderly or for disaster relief. To get there, we must build robots that can do a wide variety of tasks and adapt in real-time to the complexities of the real world. Our work is one step in this direction, but a lot of future research remains.

What is OpenAI working on next? Any exciting developments we should know about?

Peter: We will continue our work toward general-purpose robots. We’re excited to explore more tasks for robot hands beyond Rubik’s Cube.

Toolbox Perspective:

At a time when robots have invaded the factory floors, OpenAI’s latest research can take it closer to implement it in real-world setup. Though it is not the first time a robot has solved the Rubik’s Cube, this OpenAI project can be pegged as an achievement since the researchers didn’t program each movement into the robotic hand, thereby saving several years’ worth of time.

Instead, the researchers succeeded in building a system that learned to solve the Rubik’s Cube mainly on its own. OpenAI’s research centers around reinforcement learning and the latest research takes it closer to the broader goal of making robots more adaptable for real-world environment.

About OpenAIOpens a new window :

OpenAI is an AI research lab with the mission to ensure that artificial general intelligence benefits all of humanity.

Found these views on OpenAI’s AI latest robot interesting? Follow us on TwitterOpens a new window , Facebook, Opens a new window and LinkedInOpens a new window to get the latest news and updates from the world of technology.