NVIDIA's researchers have created a new artificial intelligence (AI) technique meant to train robots to perform a task by first analyzing and learning from human actions before mimicking them. The researchers, led by Stan Birchfield and Jonathan Tremblay, used NVIDIA's TITAN X GPUs to teach a series of neural networks that incorporated perception, program generation, and execution.
The goal of the new method is to allow robots to interact with humans in a workplace and improve how they communicate with one another in real-world environments. To train neural networks, the new AI technique uses synthetic data so that robots can simulate a human demonstration by generating a description of the steps needed to perform the task on their own. The description is readable to humans as there might be a need later to check whether the robot accurately interpreted the human action and make the necessary corrections before the robots execute a particular task. The researchers also used an image-centric domain randomization method on a robot for the first time in order to generate synthetic data with infinite amounts of variation. The AI technique is demonstrated on a number of colored blocks and a toy car, as shown in the video below in which the system learns about the physical relationship of blocks regardless of their position. After seeing the stacks of cubes, the robot was able to set the cubes in the proper place by inferring an appropriate program based on its interpretation of the block arrangements.
NVIDIA's new AI technique will be presented at the International Conference on Robotics and Automation, which will take place in Brisbane, Australia from today through Friday. It is just one of the many AI-related initiatives the company has been investing in recent times as part of its AI push. Last March, NVIDIA formed a partnership with ARM to integrate deep learning inferencing into a huge number of mobile and Internet of Things (IoT) devices. The Delaware-based tech company also teamed up with Baidu last July to collaborate on AI-powered technologies across different markets including cloud computing, self-driving cars, and digital home assistants.