Google Robot teaches itself to walk without human help in just a few hours
In a conversation about A.I., there is always one question that gets asked, and that is, are machines intelligent. In our AI class we have had multiple debates and conversations about this particular issue, but we never really got to a point we all agree on. So are machines intelligent? If so what is intelligence? How do we define it? And how do we test it? Tests such as the Turing test and the Chinese room have used approaches such as conversation to test that and help the AI world define that. So what is next?
Well Google believes they have achieved the first step in making or distinguishing between what it means to be an intelligent machine. Google researchers have built a four-legged robot capable of learning how to walk without any human interference. Google's Robotics Team and the Georgia Institute of Technology used a technique called deep reinforcement learning, which was programmed to perform the task of learning how to walk.
There is more. Not only did this four-legged robot learn how to walk, it learnt on three different terrains: Flat ground, a doormat with crevices and a soft mattress.
"Our system can learn to walk on these terrains in just a few hours, with minimal human effort, and acquire distinct and specialized gaits for each one," states a paper explaining how the system works.
The robot started by rocking backwards and forwards, before using trial and error to understand that it can propel itself forward by bending its legs in the correct sequence. And so on top of that, another algorithm was used to help the robot stand back up whenever it fell over. So, using these algorithms the robot was able to walk across all the three terrains without falling over, within a few hours. Which is quite an achievement, I believe. Once it had learned to walk, the researchers were able to hook up a game controller and move it using the moves the robot learnt.
A Business Insider article stated that ,“The main benefit of a robot not requiring manual input is that the AI framework can learn how to walk on a huge range of surfaces without the need to program each necessary gait individually.”
What Google researchers hope to do now is that, to test it on robots in different types of situations, which in turn could be developed to be used in real life situations. This is not the first time Reinforcement learning has previously been used. It has been used on robots such as video games and Jenga.
Another fun fact about reinforcement learning is “Last year, a team from the Computational Robotics Lab at ETH Zurich demonstrated a 3D-printed robot capable of autonomously teaching itself how to ice skate.” Which I think is fascinating, and you can read more about it here.
This would be an example of a robot that couldn't pass the Turing test but might still be considered intelligent because of its ability to learn how to do things without having to be taught. As AI becomes more advanced it will be interesting to see how the definition of an intelligent machine changes since it is not easily defined right now.
ReplyDeleteI agree. This A.I. is displaying a different kind of intelligence - the kind associated with children. Far more than responding in human-like ways, I think that robots that can learn from their mistakes and teach themselves new things based on what they already know is a better measure, since children learn how to walk and talk through repeated effort, and the rate at which they develop these skills shows their potential for intelligence.
DeleteIt's really cool how they mentioned that two individual robots have unique gaits. It totally makes sense though, considering they went through their own processes to learn how to walk. It just kind of drives the point that they figured it out on their own.
ReplyDeleteThe fact that robots can do this today shows that robots don't necessarily have to pass the Turing test as Maegan pointed out to be considered intelligent. Reinforcement learning is definitely making changes in the world of AI and how we see it.
ReplyDeleteI think that the Turing Test is extremely specific to how Alan Turing envisioned our communication and connection to AI considering it is mainly based around verbal communication. Anything in this world can be intelligent without the ability to verbally communicate. So in this example, the robot would be unable to hold a conversation yet has the ability to learn to walk on its own, which could be considered intelligence. I think there needs to be new standards to determining whether a robot is intelligent or not.
ReplyDeleteI agree with Sam and think that there needs to be a new standard to determine intelligence. The Turning test tests intelligence based on human standards yet AI will struggle to achieve that. It's similar to the Einstein quote the says "Everybody is a Genius. But If You Judge a Fish by Its Ability to Climb a Tree, It Will Live Its Whole Life Believing that It is Stupid".
DeleteI agree with you about the incompetent of the Turing Test, but I don't think there's any standard to determine intelligence. Maybe we can compare this with another walking robot, but how can we compare a walking A.I. to a chess A.I.?
DeleteI think we should just scrap the concept of general intelligence. Instead, we could make scales to compare similar A.I. by their efficiency doing the tasks they are designed for.
Did the robot rely on inputs of data on how to walk? Or was it given the task of get from point A to point B with certain movements pre-programmed and it learned how to combine those movements?
ReplyDeleteThe only thing that was pre-programmed in it was for it to get up every time it fell. From the information given, i think it basically learnt the movements by it self, using reinforcement learning which is some kind of reward and punishment approach where falling would be a punishment and walking on the terrains would be the reward. That is my understanding of it.
Delete