There are a large number of reasons why non-vacuum robots are uncommon in homes. The issue of unstructured and semi-structured situations is at the top of the list.
From design to lighting to materials to human and animal occupants, no two homes are alike. Even if a robot were to successfully map every home, the spaces would constantly be changing.
This week, researchers at MIT CSAIL are presenting a novel approach to simulation-based home robot training. A portion of a person's house can be scanned using an iPhone and transferred into a simulation.
In recent decades, simulation training has become an essential component of how robots are trained. Robots may attempt and fail at tasks thousands, even millions, of times in the same amount of time it would take to do them once in the real world because of this technology.
Additionally, the repercussions of failing in a simulation are far less severe than they would be in real life. For a little while, imagine that in order to teach a robot to load a cup into a dishwasher, it had to smash one hundred actual mugs.
But in dynamic contexts such as the house, simulation is limited, much like the robots themselves. Robotics research shows that environment adaptability can be greatly increased by making simulations as easy to use as iPhone scans.
The system becomes more adaptive when something is unavoidably out of place, like moving furniture or leaving a dish on the kitchen counter. This is actually accomplished by building a sufficiently large database of environments like these.
