Teaching Real Robots: The Right Way to Set up Learning-Based Systems for Your Robot

Ali Younes, Arkady Yuschenko

Abstract:

The typical robotics system consists of perception, planning, and control modules. Each module is built upon information about its components, where modeling each part of the system plays an essential role in the design process. In practice, working with the non-linear models in robotics systems involves a lot of approximations which hinders reaching the optimal behavior for the goal
task. Alongside the difficulties in redeploying the system to solve other similar tasks. Learning-based methods provide a promising approach for robotic systems. In the last decade, the interest in incorporating machine learning into robotics systems has been evolving rapidly. The benefit of using learning is the possibility to design systems that are independent of the dynamical model of the robot, with
the flexibility to adopt new tasks and learn to excel in performance over time. The theory behind designing a learning-based system is still under development and ranges from end-to-end systems to hybrid systems that use inaccurate approximate models. In this paper, we are proposing the results of our research in learning-based systems, presenting our view on the right way to set up learning systems for
robotics. The results are a whole learning-based framework for robotics applications that work efficiently (1 h of training – 10 min robot movement) with minimum human intervention (the user has to provide video demonstrations only).