Reinforcement Learning on Robot with Variational Auto-Encoder

Abstract

Reinforcement learning enables robot to learn plentiful skills through training. The control based on end-to-end reinforcement learning output joint angle to the robot with image as input. However, using the deep network for end-to-end reinforcement learning makes it hard to converge and need much training time. And it’s also difficult to expand to other tasks because of custom-designed network. In this paper, we propose a reinforcement learning structure with variational auto-encoder that can be applied to different goals and reduce training time. Firstly the auto-encoder is trained with images that capture random robot actions. Then the reinforcement learning network is trained with the latent space vectors from auto-encoder instead of raw image. After finish training, the robot can reach a state similar to the state in expected image that we input.