23.05.2019, by Imke van der Sanden
We have seen them in science fiction movies; robots that are so human-like they are able to do our tasks for us. However, the technology for humanoid robots is not as advanced as you may have come to expect from these robots in the science fiction movies. According to experts, the key issue that the industry seems to be struggling with is to close the gap between humans and robots. For instance, nothing in this world comes close to the flexibility and dexterity of the human hand. This specific challenge intrigued the engineers of OpenAI; could they figure out a way to teach a robot to manipulate an object with unprecedented dexterity with the use of artificial intelligence, reinforced learning techniques and simulations?
Robot manufacturers are working on teaching their robots a variety of tasks, without having to program them for each specific task and by using the same algorithm as their video-game playing bot OpenAI Five. The task OpenAI came up with was to train a human-like robot hand to manipulate a six-sided cube, by moving it from one position to another so one particular side was facing up. The system, called Dactyl, uses movements similar to those used by real people to rotate a block into any direction. The team of OpenAI truly made a breakthrough when they started messing with the simulation. They added random visual noise and changed the colors of the virtual hand and cube, but it did not stop there. The team randomised the size of the cube, its weight and how slippery its surfaces were, all in order to give the AI a better understanding of what it might be like to manipulate the cube in the real world. Even though the simulation is not 100% true to life, it dealt with enough variations for the system to learn how to deal with the unexpected. That is how Dactyl autonomously taught itself three different dexterous manipulation behaviours in order to reach its goal, even while the simulation was being messed with.
Once fully trained, Dactyl was able to reposition the cube up to 50 times in a row without dropping it. As of right now Dactyl has taught itself three human-like strategies to achieve its goal: finger pivoting, sliding and finger gaiting. These strategies were not taught to Dactyl by human instructions, but emerged as behaviours from its training of trial and error, for decades at a time.
A breakthrough in robotic manipulation?
Experts in the AI and robotics industry praise OpenAI’s work in The Verge, but are cautious in calling it a breakthrough in robotic manipulation. According to these experts, the results are still limited to a specific task (rotating an object of convenient size) in favourable conditions (where the hand is facing up, making it harder for the object to be dropped than if the hand was positioned in a different way).
So what does this mean for our future? Personally, I do not believe that we need to fear dexterous robots from ‘taking over the world’, anytime soon. Dexterity in robots is something that needs to be explored on a broader level, by taking more variables into account. As mentioned at the beginning of this article, the difference between robots and humans is something the robotic industry still struggles with. It is a time consuming and complicated process for robots to learn simple tasks. For instance, tasks that took Dactyl over a 100 years of learning can be picked up by a human with only very little trial. However, with the way Dactyl was autonomously able to teach itself human-like strategies, it is safe to say that dexterous robots, with the help of artificial intelligence, are most definitely trying to catch up. Perhaps dexterous robots are not taking over the world just yet, but I believe that we can expect them to assist and compliment our work in certain industries in the upcoming years.
Want to see Dactyl at work? Watch this short video from OpenAI for more information about Dactyl and some visuals.