Google designs a robot that uses an advanced language model

Google seeks to combine robotics and the ability to converse chatbots to develop robots capable of interpreting the orders given and evaluating possible actions based on their abilities.

Google’s research division has announced that it has integrated the PaLM language model with robots designed by Everyday Robots, a subsidiary of Alphabet, in a new system dubbed “PaLM-SayCan”.

“Today, robots largely exist in industrial settings and are painstakingly coded for specific tasks. It is therefore impossible for them to adapt to the unpredictability of the real world. That’s why Google Research and Everyday Robots are working together to combine the best of language models with robot learning,” Vincent Vanhoucke, Senior Director of Robotics Research at Google, presented in a blog post earlier this month. week.

Moravec’s paradox

The PaLM is a model with 540 billion parameters that uses a new technology to coordinate thousands of chips, known as Pathways, also invented by Google.

He draws his understanding of the world from Wikipedia, social media and other web pages. Similar artificial intelligence is the basis of chatbots or virtual assistants, but it has never been applied to robots so extensively before, Google points out.

This makes it possible to “communicate with household robots through text or speech” and improve “the robot’s overall performance and its ability to perform more complex and abstract tasks by exploiting the knowledge of the world encoded in the model. of language,” the company describes.

The research thus tackles the “Moravec paradox”, i.e. “the idea that in robotics, it is the easiest things that are the most difficult for a robot to program”, notes Vincent Vanhoucke.

Long-term task planning

The system is still in its infancy, but Google is seeing some progress.

“When the system was integrated with PaLM, compared to a less powerful base model, we saw a 14% improvement in the planning success rate, or ability to define a workable approach to a task. We also saw a 13% improvement in execution success rate, or the ability to complete a task. This represents half of the number of planning errors made by the basic method”, points out Vincent Vanhoucke.

He adds that “the most significant improvement, 26%, is in planning for long-term tasks, that is, those that have eight or more steps.”

Leave a Comment