The International Conference on Robotics and Automation (ICRA), held in Brisbane, Australia, had many surprises in store for attendees. One of them was the unveiling of research conducted by Nvidia on training robots by imitating processes performed by humans.
Until now, robots have been trained to perform a series of actions for a set number of repetitions, in an environment isolated from the humans who programmed them. In the demonstration, Nvidia used a few solid cubes of different colors and stacked them in a fixed pattern to teach the robot how to reproduce that particular action.
The sequence of actions was successful and the system developed an autonomous program to emulate the process followed.
The video released by Nvidia shows the system being trained using the Nvidia Titan X GPU, and learning from a single demonstration performed by a human.
The research aims to reduce the cost of reprogramming a robot to perform a set of tasks. “What we’re interested in doing is making it easier for a non-experienced user to teach a robot a new task by simply showing it what to do,” said Stan Birchfield, the lead scientist involved in the research.
With this research, Nvidia has joined the group of companies such as Google and SRI, which are making advances in the field of Artificial Intelligence aiming to develop AI-powered systems that are intelligent and have "common sense".
