Luckeciano M. created project Deep Reinforcement Learning for Humanoid Robot Walking and Kicking

### Deep Reinforcement Learning for Humanoid Robot Walking and Kicking

In this project, we apply modern Deep RL algorithms for optimizing Walking Engine and Kicking in Soccer 3D Simulation. The parameter optimization of walking and kicking is a big challenge in Robocup Soccer 3D Simulation enviroment. Until last year, the walking and kicking parameters are optimized by CMA-ES algorithm (a variant of genetic algorithms). This algorithm does not scale for hundreds or thousands of parameters, and it's proved that the bigger the number of parameters, the better is kicking and walking. Now, the Soccer 3D community started to apply Neural Networks for model the kick engine, and the results are very impressive. Due this, we started a project for finding better architectures for model the kick and started to apply the same techniques in walking engine. We propose to learn the entire walking engine with these techniques, replacing the traditional Double Inverted Pendulum-based walking engine.