Shulou(Shulou.com)11/24 Report--

CTOnews.com, May 14 (Xinhua) A team of researchers from Princeton University, Stanford University and Google have developed a robot called TidyBot that can understand English instructions and perform housework using OpenAI's GPT-3 Davinci model. The robot can automatically complete tasks such as sorting and washing clothes, picking up rubbish on the ground and picking up toys according to users' preferences.

The GPT-3 Davinci model is a deep learning model that is part of the GPT model series and can understand and generate natural languages. This model has a strong ability to summarize, and can learn complex object attributes and relationships from a large number of text data. The researchers used this ability to have the robot position according to several sample objects provided by the user, such as "yellow shirt in the drawer, dark purple shirt in the wardrobe, and white socks in the drawer." then let the model summarize the general rules of user preferences and apply them to future interactions.

"our basic insight is that the summarization ability of LLM (large language models) matches the generalization needs of personalized robots," the researchers wrote in the paper. LLM demonstrates the amazing ability to generalize through summarization, leveraging complex object attributes and relationships learned from massive text data sets. "

They also wrote: "unlike traditional methods that require expensive data collection and model training, we show that LLM can generalize the robot field directly out of the box, taking advantage of their powerful summarization ability learned from massive text data."

On the paper website, the researchers showed a robot that could divide laundry into light and dark colors, recycle beverage cans, throw away trash, pack bags and cutlery, put scattered items back in place, and put toys in drawers.

The researchers first tested a text-based benchmark dataset in which user preferences were entered and asked the model to create personalized rules to determine the attribution of items. The model summarizes the examples into general rules and uses summaries to determine where new items are placed. The base scene is defined in four rooms, with 24 scenes in each room. Each scene contains two to five places to put items, and there are the same number of seen and unseen items for model classification. The test achieved 91.2% accuracy on unseen items, they wrote.

When they applied this method to the real-world robot TidyBot, they found that it could successfully clean up 85% of the objects. TidyBot tested in eight real scenarios, each with a set of ten items, and ran the robot three times in each scene. According to CTOnews.com, in addition to LLM,TidyBot, an image classifier called CLIP and an object detector called OWL-ViT are used.

Danfei Xu, an assistant professor at the School of Interactive Computing at Georgia Tech, said of Google's PaLM-E model that LLM gives robots more problem-solving capabilities. " Previous mission planning systems mostly rely on some forms of search or optimization algorithms, which are not very flexible and difficult to build. LLM and multimodal LLM enable these systems to benefit from Internet-scale data and can be easily used to solve new problems. " He said.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.