Stochastic Neural Control using Raw Pointcloud Data and Building Information Models

Recently, there has been a lot of excitement surrounding the use of reinforcement learning for robot control and navigation. However, many of these algorithms encounter difficulty navigating long or complex trajectories. This paper presents a new mobile robot control system called Stochastic Neural Control (SNC), that uses a stochastic policy gradient algorithm for local control and a modified probabilistic roadmap planner for global motion planning. In SNC, each mobile robot control decision is conditioned on observations from the robot sensors as well as pointcloud data, allowing the robot to safely operate within geometrically complex environments. SNC is tested on a number of challenging navigation tasks and learns advanced policies for navigation, collision-avoidance and fall-prevention. Three variants of the SNC system are evaluated against a conventional motion planning baseline. SNC outperforms the baseline and four other similar RL navigation systems in many of the trials. Finally, we present a strategy for transferring SNC from a simulated environment to a real robot. We empirically show that the SNC system exhibits good policies for mobile robot navigation when controlling a real mobile robot.