PALM-SayCan, artificial intelligence for home help robots
It is very important that robots designed to help in the home can correctly understand and interpret human commands, and for this large-scale language models are being used to prevent inappropriate or dangerous actions from being carried out.
At the moment the risk is not very high, but imagine the scene in the near future: a human tells a robot to help him turn off an appliance, and the robot decides to break it to finish earlier, something that could happen in some languages when the verb to turn off can have more flexible meanings than in Spanish.
The fact is that the Everyday Robots researchers are working on the issue together with Google Research. Both companies are owned by Alphabet. For this they have created ‘SayCan’, language models based on real-world skills previously trained to develop the PaLM model, or Pathways Language Model. Combining them, we have PaLM-SayCana way to simplify communications between humans and robots.
The fundamental thing is to make the robots understand the context of the question that a human asks, and that they respond according to the environment.
So far PALM-SayCan has been tested with a robot in a kitchen environment. When the researcher asks “I spilled my drink, can you help me?”, the robot doesn’t suggest anything, it just comes back with a sponge and learns by experience. With traditional models, the robot could respond with things like “a vacuum cleaner might be helpful,” without taking any action, since “help” can mean many different things.
There is an article, titled “Do what I could do, not what I say”, where Google explains how they structure the robot’s planning capabilities to identify one of its ‘skills’ based on a high-level instruction from a human being , and then evaluate the probability that each possible skill is for the fulfillment of the instruction.
When they receive an instruction from a human, SayCan combines the probabilities of a language model with the probabilities of performing an action, and this is repeated until the process is successful.
