Toyota Research Institute shares its efforts to train robots using a huge collection of behavioral models. They are trying to train robots to learn from many differe actions and behaviors.
According to Hoshio, learning may be the most exciting froier in all of robotics. This field dates back several decades. For example, the 1980s brought exciting advances in visual learning, but a number of research projects at schools like CMU, MIT, and UC Berkeley poi to a future in which robots learn much like their human couerparts. .
At the TechCrunch Disrupt eve, Toyota Research Institute (TRI) will showcase advances in research that can literally teach a robot a new skill overnight.
Gil Pratt, TRI’s CEO and chief scieist, highlighted the remarkable speed of the new technology it’s working with, saying, “In the past, machine learning required millions of training items to work properly, but now it seems that only dozens of cases are needed to train these models. This developme allows processes to become faster and more efficie, even in cases where diversity in education is less importa.”
The system developed by TRI combines differe techniques to teach robots differe skills. The research branch of this car manufacturer says that using this method, it has trained robots with 60 skills, but the existing models are not able to solve the problem alone.
“We’ve seen great progress with the emergence of large language models that use them to convey this high level of cognitive ielligence to robots,” said TRI senior research scieist Benjamin Burchfill. If you have a robot that picks up something, now instead of specifying an object, you can tell it to pick up a Coke can, or you can tell it to pick up a shiny object. That’s really cool, but still, these models have limitations and can’t do things like plug in a USB device or pick up paper towels. They are really helpful, but they don’t solve that part of the problem. We are focused on addressing this limitation and are very excited about the progress we have made so far.”
One of the main advaages of this method is the ability to program skills that enable robots to work successfully in differe types of environmes. This is importa because bots typically struggle when working in less structured or unorganized environmes. For example, it is easier for a robot to operate in a warehouse compared to a road or house because warehouses are usually designed with a fixed layout and fewer obstacles.
Ideally, you wa a robot that can adapt to unexpected situations. TRI’s main focus is on developing systems that can help older people live independely, which is especially importa in places with an aging population like Japan. Based on this, the goal is to design a versatile system that can work in differe environmes and can seamlessly adapt to changes.
