An AI Robot Learnt How to Solve A Rubik's Cube One-Handed

• 
  via Giphy

  Artificial intelligence research organization OpenAI is on the quest to build a general purpose, jack of all trades kind of robot that can self-learn. In other words, they want a robot to build a robot that can do all the things a human can do without being programmed to do them. 

  Brushing your teeth, making origami, and writing with a pen might seem like effortless tasks, but we tend to forget that our species has spent hundreds of millions of years evolving to get to where we are today. It's actually quite difficult to build a robot that can bypass this evolutionary acquirement and handle things like a human hand can. 

• Advertisement
• 
  via GZT

  There are many robots out there that can solve Rubik's cubes much faster than Dactyl (this one can do it in less than a second). But any old robot can be programmed to solve a specific task. That's easy. The thing about Dactyl is that it is a self-learning robot, like us humans. We have to do things a number of times before we learn how to do them properly. And robots that are programmed to do one specific task are limited to that task - they can't learn new skills. 

  
  

  
  

• 
  via Giphy

  OpenAI didn't exactly program Dactyl to solve the Rubik's cube, but rather gave it an end goal of solving a scrambled cube using a type of incentive-based deep learning called reinforcement learning. This way, the robot figured out the task on it's own rather than doing what it was told - or programmed - to do. 

  This kind of AI is a lot easier to develop virtually, because the trial and error process can be sped up in a virtual world, making thousands of years of learning pass in a few days. The same can't be done with real world, physical bots like Dactyl. So, naturally, it's a slow process to get Dactyl acting like a human hand. 

• 
  via Giphy

  The researchers at OpenAI don't have thousands of years to monitor Dactyl as it learns by trial and error, so they created thousands of different simulated environments where a virtual Dactyl learnt to solve the Rubik's cube. After the robot had had thousands of virtual years of trial and error with a Rubik's cube, it was attempted in real life. 

  
  

  
  

• Advertisement
• 
  via Giphy

  After a few days of training (which translates to thousands of years in the virtual world), even with impediments on it that it had not been trained for like rubber bands constricting the fingers and confetti being dropped on it, Dactyl was able to solve the Rubik's cube - albeit very awkwardly and slowly. 

  However, in their inspirational promotional video, OpenAI conveniently forgot to mention that Dactyl dropped the cube 8 out of 10 times. This goes to show that although the robot can solve a Rubik's cube through self-learning, which is an achievement, we're still a very long way off from living in a world where humanoid robots make us breakfast, clean the house and take the kids to school. Until then, we'll have to be satisfied with the robots we already have in our lives. 

  
  

Tags