Addressing Reward Engineering for Deep Reinforcement Learning on Multi-stage Task

In the field of robotics, it is a challenge to deal with multi-stage tasks based on Deep reinforcement learning (Deep RL). Previous researches have shown manually shaping a reward function could easily result in sub-optimal performance, hence choosing a sparse reward is a natural and sensible decision in many cases. However, it is rare for the agent to explore a non-zero reward with the increase of the horizon under the sparse reward, which makes it difficult to learn an agent to deal with multi-stage task. In this paper, we aim to develop a Deep RL based policy through fully utilizing the demonstrations to address this problem. We use the learned policy to solve some difficult multi-stage tasks, such as picking-and-place, stacking blocks, and achieve good results. A video of our experiments can be found at: https://youtu.be/6BulNjqDg3I.