Artificial Intelligence (AI) is all about the recreation of human intelligence or our thought process, natural skills and inscribing them into machines, so-called robots. The recreated skills maybe speaking different languages, listening, reacting and more. This can be done by the machine by analysis of the ongoing situation and react accordingly using sensors or any human input. For robots to perform a task in the real world, the task must be communicated to it clearly to achieve the desired result and also the robot must perform the task robustly according to the changes in the real world.
Relating to Artificial Intelligence, NVIDIA has demonstrated a robot, that learns and completes tasks just by observing real-world demonstrations. They claim this development to be ‘first of its kind’. A team led by Stan Birchfield and Jonathan Tremblay has demonstrated about a system of synthetically trained neural networks that contained in a human-readable program for Human-Readable plans from real-world demonstrations to perform tasks associated with perception, program generation and program execution using a graphics card.
For perception, they used image-centric domain randomization leveraging convolutional pose machines, which can be applied to any camera without any assumptions about camera pose. For program generation and execution, they used fully connected networks which are simple but efficient. And this work drew inspiration from ‘one shot imitation learning’, which is commonly applied in different tasks, in isolation.
They say that-“For robots to perform useful tasks in real-world settings, it must be easy to communicate the task to the robot; this includes both the desired result and any hints as to the best means to achieve that result.
With demonstrations, a user can communicate a task to the robot and provide clues as to how to best perform the task.”