ICRA 2019 Video Supplementary Material This video explains the intuitions behind RMP and a RMP model designed for an Ackermann steering vehicle. It compares RMP with direct prediction of laser scan and control command in the Gibson simulation environment. Finally it demonstrates real-world navigation and obstacle avoidance behavior.

A Complete Navigation Trace This video shows a complete navigation trace using our neural RMP model. The input to the model is a 224x224 color image shown in the top right corner. A simple A* planner provides waypoints (shown as blue circle). Note that the planner only provides a greedy path so naively following it would cause collisions. Our neural RMP model takes the surrounding obstacles into consideration and hence only uses the waypoint as a hint and does not follow it blindly. The control commands (velocity and steering angle) are computed from the kinematic model of the vehicle. Though not shown here, our neural RMP model can naturally handle dynamic obstacles observed by the camera as shown in the first video.