Human Language Accelerates Robotic Learning

Researchers in Princeton have found that human language descriptions of devices can accelerate learning from simulated robotic arms that can lift and use tools.

New research supports the idea that AI training can make autonomous robots more adaptable to new situations, improving their efficiency and safety.

By adding a description of the form and function of a device to the robot training process, the robot’s ability to design new devices is enhanced.

ATLA approach for training

The new method is called Accelerated Learning of Tool Manipulation with Language or ATLA.

Anirudha Majumdar is an Assistant Professor of Mechanical and Aerospace Engineering at Princeton and Head of the Intelligent Robot Motion Lab.

“More information in the form of language can help robots learn to use devices faster,” Majumdar said.

Also Read :  Deep Learning Algorithm Can Hear Alcohol in Voice

The team questioned the GPT-3 language model to get a device description. After trying various stimuli, they decided to use “Describe the [feature] Of [tool] In a detailed and scientific response ”with the features, appearance or purpose of the device.

Karthik Narasimhan is an assistant professor of computer science and author of the study. Narasimhan is also a faculty member in Princeton Natural Language Process (NLP) and contributed to the original GPT language model as a research scientist at OpenAI.

Narasimhan said: “Because these language models are trained on the Internet in some sense, you may think that it is a different way of extracting information that is more efficient and comprehensive than using crowdfunding sources or specific websites for Device Description “.

Also Read :  Adept AI Labs discusses new foundation model at Oracle CloudWorld

Simulated robot study experiment

The team selected 27 training tools for their simulated robot learning experiment with tools ranging from axes to baskets. The robotic arm is assigned four different tasks: slide the lifting device, use it to push the cylinder through the table or a wooden hammer into the hole.

The team then developed a set of policies using machine learning methods, with and without language information. Policy implementation was compared on separate tests of nine devices with paired descriptions.

A method called meta-learning enhances the robot’s ability to learn with each successive task.

According to Narasimhan, the robot not only learns to use each device, but also “tries to understand the descriptions of each of the hundreds of devices, so when it sees the 101st device, it learns to use it faster.” New equipment. ”

Also Read :  82% of Marketers Say AI-Generated Content Is as Good as Human-Generated One

In most experiments, language information provided significant benefits for the robot’s ability to use new devices.

Allen Z. Ren is a Ph.D. Students in Majumdar’s team and author of research papers.

“With language training, it learns to grasp one end of the ring and use a curved surface to better control the movement of the bottle,” Ren said. “Without language, it grips the chopsticks off the curved surface and it is more difficult to control.”

“The broad goal is to get robotic systems, especially systems that are trained using machine learning – to be common to new environments,” Majumdar added.


Leave a Reply

Your email address will not be published.