Autonomous driving cars, human-like NPCs in videogames, nuclear fusion controller, faster matrix multiplication, ChatGPT and more: Reinforcement Learning (RL) is nowadays ubiquitus and when combined with sensory inputs coming from different modalities, i.e. language, image, video, sound or even nuclear physics, it is able to obtain super-human results on the task it is asked to resolve. In the last few months we have witnessed a broader recognition of such methods thanks to the huge impact they have made in our day-by-day life: Deepmind controlled a nuclear fusion at EPFL, while OpenAI released ChatGPT, a language model which interacts with the users in a conversational way. Both of these methods was trained with Reinforcement Learning, with the latter asking feedbacks from human.

In Orobix we have built a Deep RL Framework, called sheeprl, that lets anyone learn and play with agents acting in complex environments, focusing in particular on training agents to compete with human players on the MotoGP videogame. The framework has been written completely in Python, leveraging the power of PyTorch and Hydra for the deep learning modules and the configurations respectively.

This talk would mainly cover the following:

• What is Reinforcement learning, what are the main components and limitations?
• What is sheeprl, our Deep RL Framework, and what are its main components?
• How to practically make use of the framework to learn agents acting in complex environments?

The talk will firstly provide a brief introduction into the main concepts of Reinforcement Learning, what are the main components and its limitations; it will then introduce sheeprl, our Deep Reinforcement Learning Framework, and its main features; finally it will showcase how to learn intelligent agents within the framework.

You can found us at:

• https://eclecticsheep.ai/
• https://github.com/Eclectic-Sheep/sheeprl