MIT says that one day training interactive robots could be easy for anyone to do without having to program anything. Scientists are working on automating robots to give them the ability to learn new tasks by observing humans. The idea is that someday a domestic robot in a home could be shown how to do chores and then perform them for the human.
MIT has a new system that allows robots to learn complex tasks that have a lot of rules. One task during testing was setting a dinner table. The system the team developed is called “Planning with Uncertain Specifications.” MIT says that the system allows the robots to have a humanlike skill to weigh many ambiguous and potentially contradictory requirements to reach an end goal.
MIT says that the robot always chooses the most likely action to take based on a “belief” about probable specifications for the task it is supposed to do. Researchers combined a dataset with information about how eight objects could be placed on a table in various configurations. The eight objects included a mug, glass, spoon, fork, knife, dinner plate, small plate, and bowl.
The robotic arm was then tasked with setting the table in a specific configuration both in the real-world and in simulation-based on what it had seen. The test required the robot to weigh many possible placement orderings, even if items were removed, stacked, or hidden.
The robot arm made no mistakes in “several” real-world runs. Only a handful of mistakes were made in tens of thousands of simulated test runs. Specifically, the robot made six mistakes in 20,000 simulations. Scientists say that the “belief” system of the robot can be used to give the robot rewards and penalties and is built on something called linear temporal logic.