KLearn: Stochastic Optimization Applied to Simulated Robot Actions Final Report

Machine learning techniques and algorithms are prevalent in robotics, and have been used for computer vision, grasping, and legged walking. Reinforcement learning approaches have been developed over the past 15 years, with modern techniques using continuous action spaces for various robotic applications. Policy gradient learning allows various optimization techniques to quickly optimize robotic tasks, but gradient-based optimization can suffer from problems with stochastic objective functions. KLearn uses a stochastic optimization system to optimize the mean of the objective function. The optimization algorithm is applied to a simulated robot.