AI (artificial intelligence) researchers have made an important step towards the development of an ‘intelligent’ robot. If this phrase conjures up dystopian images of Arnold Schwarzenegger as a cyborg, the reality may be a far cry from what you are expecting.
Instead, the robot in question consists of a long mechanically jointed arm with a grasping hand, resembling the kind of machines often used in manufacturing. The special thing is that it has managed to teach itself, like a child, how it looks and what it can do.
Scientists have developed self-teaching robots before, demonstrating that through complex algorithms we can now design machines which can learn, using trial and error, how to complete a task to a high degree of skill. Yet in this case, researchers took on a whole new challenge by also programming the robot to discover its own self in the process. Over a time period of approximately 35 hours, this self-taught machine went from flailing randomly, with no concept of its size, shape, or abilities, to handling small objects and placing them into a container with 100 per cent accuracy.
Using these random movements, the robot was able to build up a perception of its design, which it gradually refined over the intense period of learning. Eventually, the robot managed to complete the required task despite not having been given instructions from the researchers at any point.
During this process the machine was initially allowed to measure and adjust its own movements, adapting them to the task at hand. To challenge the power of the model even further, the researchers next wanted to see what would happen if they removed even this ability – a task the scientists equate with a human finding and picking up a glass of water while blindfolded. Surprisingly, even in these demanding circumstances, the robot managed a respectable rate of 44 per cent accuracy.
Hod Lipson, who led this study at the University of Columbia, New York, explains why this research is so important. “If we want robots to become independent, to adapt quickly to scenarios unforeseen by their creators, then it’s essential that they learn to simulate themselves.” Ideally, such a machine could also respond and adapt to damage, a capacity the group tested in this study by replacing various parts of the robot with defective components. The robot managed to respond well in these scenarios, quickly identifying the faulty element and learning to work effectively despite the deformity.
Although these results are encouraging, sci-fi enthusiasts may be disappointed to hear that we are still a fair way from developing a machine with the extensive self-awareness of, say, The Terminator. Lipson has clarified that the robot’s ability to imagine itself is still primitive in comparison with that of a human. Nonetheless, this research has provided a striking demonstration of the capacity of advanced algorithms to enable a robot to learn, not just about the world around it, but also about itself.
Image: Alan Levine via Flickr