Google improves the language skills of its robots
Several technology companies are designing robots to make everyday life easier. This is the case of Alphabet, the parent company of Google, which since 2019 has been developing robots capable of performing simple tasks such as fetching drinks and cleaning surfaces. Recently, the company found a new way to help robots better interact with humans by teaching them to understand natural language commands.
To be more specific, the tech giant has combined two of its research projects, namely robotics and Artificial Intelligence (AI) language understanding. He integrated the AI called PaLM with robots from “Everyday Robots” which is one of the experimental divisions of Alphabet. And the system resulting from this combination is called “PaLM-SayCan”.
More information on “PaLM-SayCan”.
Before the integration of the “PaLM-SayCan” system, most robots only responded to short and simple instructions such as “bring me a bottle of water”. But this new technology from Alphabet provides enough real-world knowledge to help a robot interpret a wave of human command and even string together a sequence of actions to respond to it.
For example, in one test, a Google AI researcher told a “PaLM-SayCan” bot: ” I spilled my drink, can you help me? he quickly slid on his two wheels into a kitchen in an office building, spotted a sponge on the counter, picked it up with his motorized arm and brought it back to the researcher. In addition, thanks to “PaLM-SayCan” and its digital camera vision, the robot can also recognize Pepsi and Coke cans, packets of crisps and open a drawer.
A marked improvement
Although it’s still far from the perfect butler robots, the Google team has made great progress. Karol Hausman, senior researcher at Google, said that as they improve the language models, performance also improves.
Indeed, thanks to “PaLM-SayCan”, the robots were able to plan correct answers to 101 user instructions 84% of the time and executed them successfully 74% of the time. This also means that the instructions they can receive remain limited.
source
