End-to-End Probabilistic Depth Perception and 3D Obstacle Avoidance using POMDP

In most real world applications, noisy and incomplete information about the robot proximity is inevitable due to imperfections coupled with the onboard sensors. The perception and control problems go hand in hand in order to efficiently plan safe robot maneuvers. This paper proposes a method to generate robot actions directly from a sequence of depth images. The notion of Artificial Potential Field (APF) approach is used where a robot action is obtained by combining the attractive and repulsive actions generated by the goal and the obstacles respectively. This article assumes environment perception uncertainty that relates to the estimation of an obstacle’s location relative to the robot. The repulsive action generation is formulated as a Partially Observable Markov Decision Process (POMDP). A Particle Filter (PF) approach is used to estimate and track valid scene points in the robot sensing horizon from an imperfect depth image stream. The most probable candidates for an occupied region are used to generate a velocity action that minimizes the repulsive potential at each time instant. Approximately optimal solutions to the POMDP are obtained using the QMDP technique which enables us to perform computationally expensive operations prior to a robot run. Consequently, suitable repulsive actions are generated onboard the robot, each time an image is received, in a computationally feasible way. An attractive action, obtained by solving for the negative gradient of the attractive potential is finally added to the repulsive action to generate a final robot action at every time step. Lastly, the robustness and reliability of this approach is demonstrated close-loop on a quadrotor UAV equipped with a depth camera. The experiments also demonstrate that the method is very computationally efficient and can be run on a variety of platforms that have limited resources on-board.