Nvidia is Training Robots to Learn by Observing Humans


Most of the conventional industrial robots today merely repeat a well-defined task over and over again. They also perform these tasks at a safe distance from the humans that have created and programmed them.

However, most of the key players in the robotics industry are looking for new ways to make these robots work closely with the humans by learning from their movements and mimic their tasks for better performance.

The Nvidia’s new robotic lab is working on how they can make the robots and humans work in close proximity to each other so that the future robots will be able to observe humans and learn from them. The California based tech giant presented some of its on-going work that teaches robots by observing humans at the International Conference on Robotics and Automation which was held recently in Brisbane, Australia.

According to Nvidia’s senior director of robotics research Dieter Fox, their main aim is to help the next generation of robots to work closely with humans. He further added that in order to achieve this, the robots should be able to detect, track them and help them whether it is in a small-scale industry or inside a home. Fox is also professor at the University of Washington.

Today, it is easy to train a machine learning algorithm to repeat tasks and learn from its mistakes but the same thing may be very challenging when it comes to training robots that work in close proximity with humans. In order to achieve this task more efficiently, researchers at Nvidia have developed a system which will allow the robots to learn new tasks by just observing humans as they perform various tasks.

To inspect whether the robots will be able to achieve these tasks, the researchers led by Stan Birchfield and Jonathan Tremblay checked how the robots were able to stack some coloured cubes after watching humans perform the same task. While this may look like pretty simple task, it is a very vital step towards teaching new tasks to the robot.

One of the best things about this new research is that the new system also generates a human-readable description of all the steps performed by the robots during the task which will further help researchers to figure out what actually happened in case anything goes wrong. The team further plans to increase the range of tasks and the vocabulary that is needed to describe the tasks to the robots.