Mark Saroufim • 5/29/2019

Representation Theory for Robotics

This technical article discusses how to represent a robot's state efficiently to reduce memory usage and drastically speed up Reinforcement Learning (RL) training. It analyzes pixel-based representations, their limitations (like large state spaces), and optimization techniques (cropping, grayscale, frame stacking). It then introduces the need for better, model-based representations to improve sample efficiency in robotics applications.

0 comments

#computer vision #Deep Learning #Robotics