In this workshop, we'll employ Deep Reinforcement Learning to train a biped robot through simulation to walk safely and optimally following a straight line.
Designing autonomous systems (robots, vehicles, virtual assistants...) requires solving complex optimal control problems that are difficult to undertake because it's hard to define a control strategy or the objectives for each variable.
Machine Learning makes it possible to train "black box" algorithms with example data to tackle sophisticated tasks. Lying at the intersection with Game Theory, Reinforcement Learning is probably the branch with the most promising future in Automatic Control. By means of it, "agents" learn "policies" (control strategies) through trial and error. In order for these policies and the training process to be sophisticated enough, it's often useful to implement them with deep neural networks.
In our example, we'll start with a 3D physical model of the robot in Simscape Multibody™. We'll craft an accurate enough model of the environment and its rewards in Simulink. With Deep Learning Toolbox™, we'll design neural networks to codify a Reinforcement Learning algorithm and the agent's policy. Then we'll run several simulations in a parallel cluster and, once the agent is trained, we'll see how to deploy it to a programmable controller in the real robot through automatic C++ code generation