Google introduced the PaLM-SayCan, an artificial intelligence (AI) powered robot capable of responding to commands from humans. The prototype developed by Everyday Robots — a company owned by Alphabet, owner of Google — was designed to understand messages and not just perform pre-defined tasks.
According to the engineers involved in the project, the bot uses AI language processing to interpret requests and then narrows down its range of possible responses using a set of 100 general skills that it has been trained to interact with more naturally.
“You can say to the robot, “I spilled my drink, can you help?” The bot then filters that instruction through an internal list of possible actions and interprets it as “bring me the kitchen sponge,” explains prototype researcher Brian Ichter.
In the images shared by Google, it is possible to see that the robot can work quietly in a common kitchen. He is able to open drawers, handle food, separate organic from recyclable waste, and interact with his human counterparts.
According to the company, robots equipped with the PaLM-SayCan system are able to plan correct responses to 101 user instructions 84% of the time and successfully execute them 74% of the time, showing their ability to understand and do common day-to-day tasks.
“Basically, the robot hears what is being asked and obeys. They are simple commands like picking up or bringing something, but it is precisely this simplicity in the ability to understand that makes the system extremely agile and functional”, adds Everyday Robots engineer Karol Hausman.
The system behind PaLM-SayCan is based on the Pathways language model, an artificial intelligence platform that has 540 billion predefined parameters used to reproduce comprehension tasks and in the generation of natural language.
The “brain” of this system uses a combination of data in English and other languages, with information found in documents on the Internet, books, informal conversations and code stored on GitHub. This expands the robot’s vocabulary, allowing it to understand various commands.
“This model is just an experimental prototype and is not yet ready for commercial release. The execution of everyday tasks and their movements are still slow and methodical, but this is just a demonstration of everything the PaLM-SayCan system has to offer in the future”, concludes Google Research Robotics Lead Vincent Vanhoucke.