Alphabet's X wants to make robots that can learn to do mundane tasks

There has been a lot of attention paid on AI and robotics these days but the majority of those focus on either very specialized tasks or, on the other extreme, theoretical algorithms. The robots of today are still a far cry from the robots of fiction that exist not just to build cars or drive them but also to help humans in their everyday lives. To that end, Google's sibling company X has started an Everyday Robot Project to design, build, and train robots that can learn to do the messy stuff that comes as second nature to humans.

There are many things that we humans take for granted because they feel "natural" to us. These activities, like getting distracted by an impromptu water cooler chat or being unsurprised when some transient things appear and disappear every day, are, however, perplexing to robots' logical minds. For these robots to be good at assisting humans in everyday tasks, they need to be able to adapt to quickly changing elements and environments. And, for that, robots will need to learn, not programmed.

Of course, machine learning and AI are important pieces of X's Everyday Robot Project. Instead of painstakingly programming robots for one purpose and then programming them again for another purpose, the goal is to create robots that learn how to perform the tasks with minimal programming as possible. To test out its theories, X selected a task that had the right amount of difficulty while still remaining measurable in its success or failure: sorting the trash.

X's virtual robots would "practice" segregating waste in virtual offices in the cloud each night and would then apply those to real physical robots and real physical trash. The results of the "practical" activity would then be fed into the simulated training that the robots go through again that night. According to X, this had lead to a reduction of office waste contamination level to less than 5% from 20%.

It's definitely an impressive statistic but it doesn't end there. X's next big step is to train the same robots on another task without rebuilding the robot or writing a new program. It might be an impossible task, the company admits, but it will still take a shot at it. After all, it's not the Moonshot Factory for nothing.