Train the robot how to learn, make quick decisions

Adapting less images to manipulate granular materials in domain displacement

Image: The Mars rovers have groups of human experts on Earth telling them what to do. But the robots on the moon landing missions orbiting Jupiter are too far away to receive timely commands from Earth. Researchers in the Department of Aerospace Engineering and Computer Science at the University of Illinois Urbana-Champaign have developed a new learning-based method for robots on extraterrestrials to make decisions for themselves. determine where and how to take topographic samples.
see than

Credit: Grainger College of Engineering at the University of Illinois Urbana-Champaign

Mars rovers have expert groups of people on Earth telling them what to do. But the robots on the moon landing missions orbiting Saturn or Jupiter are too far away to receive timely commands from Earth. Researchers in the Department of Aerospace Engineering and Computer Science at the University of Illinois Urbana-Champaign have developed a new learning-based method for robots on extraterrestrials to make decisions for themselves. determine where and how to take topographic samples.

“Instead of simulating how to shovel every possible rock or granular material, we created a new way for the lander to automatically learn how to quickly shovel new material it encounters,” said. Pranay thangeda, a PhD. Student at the Faculty of Aerospace Engineering.

“It also learns to adapt to changing landscapes and their properties, such as the topology and composition of materials,” he said.

Using this method, Thangeda says a robot can learn to scoop up a new material with very little effort. «If it makes several unsuccessful attempts, it learns that it shouldn’t shovel in that area and it tries somewhere else.»

The proposed deep Gaussian process model is trained on an offline database with deep learning metadata with controlled implementation gaps, this model continuously divides the training set into average training and train the kernel, while also learning the multiplication parameters to minimize the residuals from the mean models. When deployed, the decision maker uses the trained model and adapts it to the data obtained online.

One of the challenges facing this research is the lack of knowledge about ocean worlds like Europa.

“Before we sent the recent rovers to Mars, the spacecraft gave us pretty good information about the terrain features,” says Thangeda. «But the best image we have of Europa has a resolution of 256 to 340 meters per pixel, which is not clear enough to identify features.»

Thangeda advisor Melkior Ornik «All we know is that Europa’s surface is ice, but it could be large chunks of ice or much smoother like snow,» said Dr. We also don’t know what’s underneath the ice.»

For some tests, the team hid the document under another layer. The robot only sees the top material and thinks it might be good to shovel. «When it actually scooped and touched the bottom layer, it knew it couldn’t scoop and moved to another area,» Thangeda said.

NASA wants to send battery-powered rather than nuclear-powered rovers to Europa because, among other mission-specific considerations, it’s important to reduce the risk of contaminating the world’s oceans with potentially hazardous materials.

“Although nuclear power supplies have a lifespan of months, batteries only have a lifespan of about 20 days. We can’t waste a few hours a day sending messages back and forth. This provides another reason why robot self-determination is so important,” Thangeda said.

This learning-to-learn method is also unique as it allows the robot to use vision and very little online experience to achieve high-quality search actions on unfamiliar terrain—significantly superior compared with non-adaptive methods and other modern methods of metaphysics.

From these 12 materials and terrain made up of a single composition of one or more materials, a database of 6,700 was created.

The team used a robot in the Department of Computer Science at Illinois. It is modeled after the arm of a lander with sensors to collect shovel data on a wide range of materials, from 1 millimeter sand grains to 8 cm rocks, as well as materials of varying volumes such as cardboard. crumbs and packaged peanuts. The resulting database in the simulation contains 100 knowledge points for each of the 67 different terrains, or a total of 6,700 points.

“To our knowledge, we are the first to open source a large-scale dataset on granular media,” said Thangeda. “We also provide code to easily access the dataset so that others can start using it in their applications.”

The model the team created will be deployed at the Jet Propulsion Laboratory of NASA’s Ocean Lander Jet Propulsion Laboratory.

“We are interested in developing autonomous robotic capabilities on extraterrestrial surfaces and especially challenging extraterrestrial surfaces,” said Ornik. “This unique method will help inform NASA’s continued interest in exploring the ocean world.

“The value of this work is adaptability and the ability to transfer knowledge or methods from Earth to an extraterrestrial body, because obviously we won’t have a lot of information before that. landing craft there. And because of the short battery life, we won’t have much time for the learning process. Landers can last for only a few days and then die, so learning and making decisions autonomously is incredibly beneficial.”

The open source dataset is available at: Drillaway.github.io/scooping-dataset.html.

Study, «Adjusting several shots for manipulation of granular materials under domain variation,» by lead authors Yifan Zhu and Pranay Thangeda, with their faculty advisors Melkior Ornik and Chris Hauserpublished in Robots: Science and Systems XIX. DOI: 10.15607/RSS.2023.XIX.048


Disclaimer: AAAS and EurekAlert! is not responsible for the accuracy of newsletters posted to EurekAlert! by contributing to the organization or for the use of any information through the EurekAlert system.

#Train #robot #learn #quick #decisions

Deja un comentario