Using robots in the industrial field has started long back and the companies are finding ways to use them more efficiently. Being harmful to humans, they are operating at safer distances from human working areas. But this results in inconvenience within the industry and hence researchers are finding ways so that humans and robots can work together without any harm to the life. At the International Conference on Robotics and Automation (ICRA) in Brisbane, Australia – Nvidia’s research team has presented its recent works of training the robots to work around by observing the humans at the new research lab in Seattle. In the conference, they showed a camera recording where a human simple comes. Which picks up the blocks and stack them. When he is done, the robots do the same by following the steps it observer.
How The Robot Works?
Nvidia’s research team is reportedly working on making a new class of robots. To train them to observe their surrounding human activities and perform the task as such. In the initial stage, the robots learn to pick up color boxes and a toy car using Baxter robot. The senior director of robotics research at Nvidia said that they want to develop the next gen of robots. Which work simultaneously with the humans in the same environment and to do that they should be able to track and detect people’s activities. Learning the behavior of human activities will help them to create robots that can work safely in human proximities without any danger.
Nvidia’s principal research scientist Stan Birchfield said that in the manufacturing environment, these robots work very well with the same trajectory executing them repeat but cannot adapt to that changes in the environment i.e. they do not learn their task rather execute the task they are defined with. It is possible to train a robot to successfully play a video game by repetition and learning. From its mistakes but the state tree will be very huge to deploy it efficiently. Instead, by observing the humans, the robots learn the new task to be performed easily.
Making The Nvidia’s Robot Work
The researchers first train the robot with a sequence of neural networks to detect the objects and find the dimensions and relation between them. Then they observe the humans and generates a program that exactly replicates the actions done by the human it observes. They show an example where the robot performs the stacking task by identifying the cubes first and then stacking them. One of the major advantages of this system is that it generates a human-readable format. The instructions and hence the errors can easily identify and correct. Some actions which are difficult to explain but relatively easy when done can use these robots to complete their tasks.
The research team uses synthetic data which obtain from a simulated environment. Training for these types of actions will be far more difficult and dangerous and such tasks do not have label data available to them. Birch said that using simulation will be a powerful paradigm to train the robots to do things which are relatively very difficult to be define. The team also admit that they have a long way to go to make such a robot fully functional. They are also trying to expand the tasks it can perform and describe the necessary vocabulary which describes these tasks.