NVIDIA trains robots capable of learning through observing people

NVIDIA has developed a way to train robots to perform actions by observing people. As a result, their robots know how to pick up and move colorful boxes and toy cars in a lab environment through a Baxter robot .

According to VentureBeat, the results of this study will be used to re-train robots and create robots that can operate safely alongside people in industrial and household environments.

This new AI system was created with the help of NVIDIA robot research lab.

"In a production environment, robots accomplish very well repetitive tasks, but they cannot adapt to changes in the work environment and do not learn their tasks" - the scientist NVIDIA's Stan Birchfield said - "Therefore, if you want to renovate a robot to perform a new job, you must bring it to a specialist to reprogram at a fairly low level, and that's a merit. What we're doing is making it easier for non-expert users to teach robots a new job just by showing it what to do. "

The robot training system that NVIDIA is referring to is a series of deep neural networks, capable of perception, planning, and control. At the same time, these networks are fully trained based on aggregated data.

"Today, we can use GPU to create an infinite amount of data that is completely free to develop and test algorithms, and its potential allows them to be used," Birchfield said. We develop robotic systems that can learn how to interact with the world around them in larger and safer ways. "

This new one was created with the help of NVIDIA robot research lab. First published at the end of last year, the lab now has 6 members and is preparing to open offices near Washington University in Seattle this summer.

The research lab will continue to work with NVIDIA's internal robotics and robotics community to explore ways to use aggregate data sets to train AI systems - NVIDIA's robot research manager Dieter Fox for know.

The results from this study can also be applied to strengthen the Isaac SDK, a framework used to train robots through the first simulation introduced in May 5/2017.

According to Fox, NVIDIA has been working in this field for a long time. For example, in the context of game making - where the main job is to set up 3D virtual environments with realistic images and content rendering, the research team not only works with all other specialized teams. private subjects, but also help them expand that expertise to make it more applicable in the robot context.

This research will be central to the creation of next-generation robots.

"We are talking about robots capable of opening doors, opening drawers, picking objects, moving them around, even interacting with people, and helping them, such as people. "The robots need to know people, see what a person wants to do, learn from them, and they must know what a person needs to help them", Fox said. .

NVIDIA is currently one of many companies (including Google and SRI International) developing environmental-sensing AI systems - which, according to Google's AI head, Jeff Dean, is "more conscious."