Mapless Navigation for Autonomous Robots: A Deep Reinforcement Learning Approach

Finding a collision-free path for mobile robots is a challenging task, especially in sceneries where obstacle information is partly observed. Our work presents a decentralized collision avoidance approach based on an innovative application of deep reinforcement learning. The approach takes the spare 10-dimensional range findings and the target position in mobile robot coordinate frame as input and the continuous action commands as output. Traditional method for finding collision-free paths deeply depends on extremely precise laser sensor and the map making work of the roadblocks is inevitable. Our work shows that, using an asynchronous deep reinforcement learning method, a mapless path planer can be trained from start to finish without any manual operations. The trainer is available in other virtual environment directly. We compare a traditional method with the asynchronous method and find that our asynchronous method can decrease training time at beginning.