Exploring a new way to teach robots, Princeton University researchers have discovered that human language descriptions of tools can speed learning to simulate raising a robotic arm and using a variety of tools.
The findings build on evidence that providing richer information while training artificial intelligence (AI) can make autonomous robots more adaptable to new situations, improving their safety and effectiveness.
Adding descriptions of the tool’s shape and function to the training process for the robot improved the robot’s ability to handle newly encountered tools that were not present in the original training set. A team of mechanical engineers and computer scientists presented the new method, Accelerated Tool Manipulation Learning with LAnguage, or ATLA, at the Robot Learning Conference on December 14.
Robotic arms have great potential to help with repetitive or difficult tasks, but training robots to handle tools effectively is difficult: tools have a wide variety of shapes, and a robot’s dexterity and vision are out of step with a human.
“Additional information in the form of language can help a robot learn to use tools faster,” said study co-author Aniruddha Majumdar, an assistant professor of mechanical and aerospace engineering at Princeton who leads the Intelligent Robotics Laboratory.
The team obtained the tool descriptions by querying GPT-3, a large language model released by OpenAI in 2020 that uses a form of artificial intelligence called deep learning to generate text in response to a prompt. After trying the various prompts, they settled on using “Describe [feature] From [tool] in a detailed and scientific response”, where the attribute was the form or purpose of the tool.
“Because these language models are trained online, you can somehow think of this as a different way of retrieving that information,” Karthik Narasimhan said, in a way that is more efficient and comprehensive than using crowdsourcing or deleting specific websites to describe the tools. Assistant professor of computer science and co-author of the study. Narasimhan is a Principal Faculty Member of the Natural Language Processing (NLP) group at Princeton, and has contributed to the original GPT language model as a Visiting Research Scientist at OpenAI.
This work is the first collaboration between the research groups of Narasimhan and Majumdar. Majumdar is focused on developing AI-based policies to help robots — including flying and walking robots — generalize their functionality to new settings, and was interested in the potential of “the recent huge advances in natural language processing” to benefit robot learning, he said.
For their simulated robot learning experiments, the team chose a training set of 27 tools, ranging from an ax to a mop. They gave the robotic arm four different tasks: push the tool, lift the tool, use it to sweep a cylinder along a table, or drive a stake into a hole. The researchers developed a set of policies using machine learning training approaches with and without language information, and then compared the performance of the policies on a separate test set of nine tools with associated descriptions.
This approach is known as meta-learning, in which the robot improves its learning ability with each successive task. It’s not just about learning to use each tool, Narasimhan said, but also “trying to learn how to understand the descriptions of each of these hundreds of different tools, so when he sees Tool #101 he’s quicker in learning to use the new tool.” “We’re doing two things: we’re teaching the robot how to use tools, but we’re also teaching it English.”
The researchers measured the robot’s success in pushing, lifting, sweeping, and knocking with the nine test tools, and compared the results achieved with policies that used language in the machine learning process with those that did not use language information. In most cases, the linguistic information provided significant advantages to the robot’s ability to use new tools.
Allen Z. said: Student at Majumdar Group and lead author of the research paper.
“Through language training, he learns to grip the long end of the crowbar and use the curved surface to better constrain the movement of the bottle,” said Ren. “Without the language, the crowbar held close to the curved surface and was difficult to control.”
The research was supported in part by the Toyota Research Institute (TRI), and is part of a larger TRI-funded project in Majumdar’s research group that aims to improve the robots’ ability to operate in novel situations that differ from their training environments.
“The overall goal is to get robotic systems — specifically, those that are trained using machine learning — to generalize to new environments,” Majumdar said. Other work supported by his TRI team addressed failure prediction for vision-based robot control, and he used a “hostile environment generation” approach to help robot policies work better in conditions outside of their initial training.
Article, Utilize language to quickly learn to manipulate toolsIt was presented on December 14 at the Machine Learning Conference. Besides Majumdar, Narasimhan, and Ren, co-authors include Bharat Govil, Princeton class of 2022, and Tsung-Yin Yang, who completed his Ph.D. in electrical engineering at Princeton University this year, and is now a machine learning scientist at Meta Platforms Inc.
In addition to TRI, support for the research was provided by the US National Science Foundation, the Office of Naval Research, and Princeton University’s College of Engineering and Applied Sciences through William Addy ’82 generosity.
#Words #prove #educational #tools #robots #ScienceDaily