You could one day teach a robot just by executing a task yourself.
Till now, teaching a robot to accomplish a task has usually include either trial-and-error, direct coding, tests or handholding the machine. Shortly, however, you might just have to execute that task like you would any other day. MIT scientists have made a system, Planning with Uncertain Specifications (PUnS), that assists bots learn complex tasks when they’d otherwise misstep, such as arranging the dinner table. Instead of the regular method where the robot gets rewards for performing the right actions, PUnS has the bot hold “beliefs” over a range of specifications and use a language (linear temporal logic) that allows it reason about what it has to do right presently and in the future.
To move the robot toward the right outcome, the team set standards that helps the robot meet its overall beliefs. The standards can satisfy the formulas with the highest probability, the greatest number of formulas or even those with the minimum chance of failure. A designer could improve a robot for safety if it’s working with dangerous materials, or steady quality if it’s a factory model.
MIT’s system is much more efficient than traditional approaches in early testing. A PUnS-based robot made just six mistakes in 20,000 attempts at arranging the table, even when the researchers threw in difficulties like hiding a fork — the automaton just completed the rest of the tasks and came back to the fork when it appeared. In that way, it showed a human-like ability to set a clear overall aim and create.
The developers eventually want the system to not only learn by watching, but response to feedback. You could give it verbal improvements or a review of its performance, for instance. That will involve much more work, but it indicates at a future where your household robots could adjust to new duties by watching you set a pattern.