Model Predictive Control with a Learned Model for Safe Exploration

Modern information technologies and the advent of machines powered by Artificial Intelligence (AI) have already strongly influenced the world of work in the 21st century. Advances in Artificial Intelligence (AI) technology and related fields have opened up new markets and new opportunities for progress in critical areas such as health, education, energy, economic inclusion, social welfare, and the environment. In recent years, machines have surpassed humans in the performance of certain tasks related to intelligence. Although it is unlikely that machines will exhibit broadly-applicable intelligence comparable to or exceeding that of humans in the next 20 years, it is to be expected that machines will continue to reach and exceed human performance on more and more tasks. Through the use of artificial intelligence, robots will be able to independently assess what is happening around them and make decisions on the actions they need to take. For this reason, the ability to robustly guarantee safety becomes absolutely essential: for instance, consider tasks such as steering an autonomous vehicle along a given reference track, precise machine tooling, or control of any autonomous system that interacts with humans. All these applications have in common that the autonomous system should be able to complete the task without harming itself and its surroundings since in real contexts safety is a critical factor and errors are not acceptable. This Thesis studies the Safe Exploration Problem without assumption about prior knowledge of the system dynamics and the constraint function. To cope with this problem we combine a medium-size Deep Neural Network model with Model Predictive Control to achieve excellent sample complexity and sample efficiency in order to produce plausible gaits to accomplish locomotion tasks. The Model Predictive Control uses the random shooting methods to optimize the control sequence considering the dynamics of the model and the constraints learned by the Deep Neural Networks. We evaluate the performance of our Robot in the Safety Gym environment that provides a set of tools for safe exploration research.