In this paper, a deep reinforcement learning approach was implemented to achieve autonomous collision avoidance. A transfer reinforcement learning approach (TRL) was proposed by introducing two concepts: transfer belief — how much confidence the agent puts in the expert’s experience, and transfer period — how long the agent’s decision is influenced by the expert’s experience. Various case studies have been conducted on transfer from a simple task — single static obstacle, to a complex task — multiple dynamic obstacles. It is found that if two tasks have low similarity, it is better to decrease initial transfer belief and keep a relatively longer transfer period, in order to reduce negative transfer and boost learning. Student agent’s learning variance grows significantly if using too short transfer period.
