Robot Hand Unexpectedly Learns Human Behaviour! - Open AI

Key Takeaways

you are watching ColdFusion TV [Music] hi welcome to another ColdFusion video as we've seen in previous episodes the field of Robotics has recently taken a leap in dexterity currently robots can run jump perform backflips and respond to real-time changes in their environment but the thing that you've probably noticed in all of these machines is that their dexterity is severely limited when it comes to handling objects with artificial hands or fingers this is one of the hardest technical challenges to overcome but that's just beginning to change engineers at the Elon Musk founded non-for-profit open AI have used the lessons learned in artificial intelligence to teach a robot how to handle objects as elegantly as a human could in this video we'll take a look so the goal for open AI was to train a robot to perform a task without being programmed to do so the specific task was to get a robot hand to manipulate a six-sided cube moving it from one position to another position so that a specific side was facing up so how would you go about doing something like this well here's the problem with current methods of using an AI to train a physical robot training is often done using reinforcement learning a method where a machine learning system acquires knowledge through trial and error the thing is reinforcement learning can take a long time usually amounting to years of experience that's fine if you want an AI to be a video game all it has to do is just play the game at a highly accelerated rate and then learn what it needs to but if you want to teach an AI a real-life task there are two problems number one you can't speed up real life no matter how much you'd like to and number two if you did want to speed up real life you'd have to create a simulation that's so accurate and realistic that it can be useful for training opened a eyes solution was to mess around with the training simulation itself they traded an accuracy for variation first in the simulation they first changed the colors of the virtual hand and cube then randomized the size and weight of the cube the friction of the surfaces how fast the hand can move and even mess with the simulation of gravity this process is called domain randomization and it proved to be a powerful tool the domain randomization gave the AI a better understanding of what it might be like for a robot hand to manipulate the cube in the real world the system had enough variations to work with that it can deal with the unexpected in a robust way the open AI team ran all of the simulations on thousands of machines in the cloud the collective knowledge was gathered and optimized before being sent back into the system for the next round of simulations the finished trained model had about a hundred years worth of experience of handling the blocks in the randomized situations [Music] some unusual things began to happen after the system was trained the robot hand became much more general in the task of handling objects for example it can rotate a wide variety of objects that has never seen before to a particular orientation when asked even more strange was the fact that the system developed human-like behaviors like sliding and finger pivoting without ever being programmed to do so [Music] open a is vision for the future is to use this method to train robots to complete a wide variety of general tasks without human input in fact open AI actually is some of the same algorithms and methods to train an AI to play complex games such as dota 2 and quake 3 in the quake 3 scenario the AI actually exceeded human level and even could collaborate seamlessly with humans within team matches and you should keep in mind that this isn't like the regular bots that you'll find in games the open AI bots initially had no idea about the game and learned how to play just from scratch the vastly different applications for these algorithms hints at the notion that general AI that can complete a wide array of tasks may be possible in terms of more direct applications creating more dexterous robots with the help of artificial intelligence could be a big push for the automation of manual labor while seeing robots in healthcare the effect that this will have on jobs is another issue entirely on a different note maybe in 10 years capable household robots may not be so far-fetched and if we look at this from a completely different angle it took a robot hundreds of years to do a sliver of what we intuitively do every day and I find that kind of interesting so what's your view on this let me know in the comments section below and that just about wraps up the video so thanks for watching this has been - go go you've been watching cold fusion if you've just stumbled across this channel feel free to subscribe and if you know someone that will be interested in this topic feel free to share this video with them alright I'll see you again soon for the next video cheers guys have a good one anything [Music] [Music] you