Scientists from MIT taught the robot to set up a table. At the same time, the device adheres to the general algorithm, but can independently change the priorities of the tasks.
MIT scientists have developed the Planning with Uncertain Specifications (PUnS) system, which helps robots master complex tasks – for example, setting the dining table. Instead of the usual method, in which the robot receives a reward for performing the right actions, PUnS teaches robots to adhere to “beliefs” on various tasks. They also pre-load instructions that allow him to reason about what he should do at a particular moment.
The MIT system is much more effective than traditional approaches to early testing. The PUnS-based robot made only six mistakes in 20 thousand test attempts to set the table. Researchers complicated his task – for example, hid a fork. After that, the robot finished laying the table, and when it found the fork, then just put it in the right place. In this way, he demonstrated the ability to set goals and improvise.
In the future, scientists want the system to not only learn by observing, but also respond to feedback. For example, she can be given corrections to tasks with the help of words and criticize her performance. This will require much more work, but scientists plan to make the robots adapt to new duties without reprogramming.