In this week, NVIDIA unveiled that its researchers have developed a unique deep learning system. This deep learning system allows a robot to learn tasks from humans by watching their actions. The main point here is the communication between the robot and human. Communication between robot and human extends to the point where the robot can watch and mimic the actions performed by the user.
• When an action is performed by a human, robot watches it. After watching a task, it develops a list of steps required to duplicate the actions performed by a human. The list of steps can be reviewed by a human to check whether they are correct before robot continues to carry out steps to mimic actions. Here is the video in which it is explained how researchers carried out a test. Researchers tried with a pair of stacked cubes, which are need to be placed in the correct order by the robot. The video explains everything.
• So, what is the advantage of a robot mimicking actions performed by a human? The advantage here is that this approach frees researchers from inputting a large amount of training data into the system. Inputting data into the computer system for instructing it to carry out a specified task is very tedious. Since ages, programming languages have been used for instructing the computer system to carry out a specified task by inputting the data into the system. But, in case of programming language, the typing becomes very tedious.
• Here, it generates the data synthetically. So, the training data required for the robot to learn and complete tasks are developed by the robot itself. For the first time, image-centric domain randomization has been tried with robots. The synthetic data produced, has a large amount of diversity and is not dependent on a camera or environmental variables. So, it requires to be once checked by a human to ensure that it will help the robot to accomplish certain task.
• This research will be presented by Stan Birchfield and Jonathan Tremblay at the International Conference on Robotics and Automation. They will now focus on how to use synthetic data to use their research and methods in different situations.
• Soon, we are going to have robots performing all actions performed by a human. This research will definitely benefit the method of instructing a computer system.