Robot Learning 2025: Foundational Models for Robotics and Scaling DeepRL - RobotLearning: Scaling PolicyGradients Part 1

l (Glen Berseth) discuss reinforcement learning (RL) in the context of robotics, focusing on policy gradients and their practical applications. Highlighting recent work on autonomous cleaning robots as an example of scaling RL in real-world scenarios. I also introduce the concept of policy gradients, explaining how they optimize a reward function without explicitly modelling the environment's dynamics. Covering different policy distributions, such as cross-entropy for discrete actions and Gaussian distributions for continuous actions, and briefly demonstrating their behaviour with an interactive tool. I include the mathematical foundation of policy gradients, emphasizing their reliance on sampling and the optimization of policy parameters to maximize expected returns. I also give a detailed explanation of a homework assignment involving a robotics dataset and using Google Colab for data processing and model training. I Emphasize the importance of proper data standardization and provide guidance on handling specific challenges like action dimension scaling. https://github.com/milarobotlearningcourse/robot_learning_2025/tree/main/hw2