Intelligent agent based low level control of complex robotic systems

Abstract

The use of intelligent agents, trained with reinforcement learning methods for control of complex mechanical systems, like humanoid robots has the potential to revolutionize the way we think about control problems. This way of learning is very similar to how we humans learn most of the things in our early age, thus proving really promising if we want to make robots able to learn tasks that require some form of intelligence. Throughout the research presented in this paper, a deep neural network based intelligent agent, with Actor-Critic architecture was trained with the Deep Deterministic Policy Gradient algorithm for the purpose of controlling a custom designed humanoid robot. For the training of the agent a simulation model of the physical robot is developed and integrated into a customizable simulated environment. The idea of low, actuator level control of complex systems by neural networks formulates the problem into a more abstract form while keeping the full control of the system without having to deal with the actual level of complexity. This can be further enhanced by expanding the abstraction from the software level to include some part of the hardware as well.