Autonomy Talks - 15/02/2021
Speaker: Dr. Diego Romeres, Principal Research Scientist, MERL, Cambridge (MA)
Title: Reinforcement Learning for Robotics
Abstract: Robot Learning has seen explosive growth and interest in recent years and it is one of the biggest modern challenges for Artificial Intelligence. In this presentation, I will describe some of the main ingredients that model-based reinforcement learning algorithms should consider for successful deployment in robotic applications. I will then propose an algorithm called Monte Carlo Probabilistic Inference for Learning COntrol (MC-PILCO). The algorithm relies on Gaussian Processes (GPs) to model the system dynamics and on a Monte Carlo approach to estimate the policy gradient during policy optimization. This defines a flexible framework in which we analyze (i) the power of combining physics first principle equations of motion with data driven methods, (ii) the structure of GPs models and (iii) how to alleviate the problem of vanishing/exploding policy gradients when using Monte Carlo methods during policy optimization. Finally, we apply MC-PILCO to robotic systems, considering in particular the case when the states are only partially measurable. We propose the idea of applying two different state observers one during model learning and one during policy optimization to better model the actual system. The effectiveness of the proposed solutions will be shown both in simulation and in real systems.
If you want to know more about Autonomy Talks, check [ Ссылка ]
Ещё видео!