A new feature selection method based on task environments for controlling robots

Abstract This paper proposes a task-based feature selection as a part of the Supervised Deep Learning (SDL) framework. The main challenge of control tasks is choosing appropriate inputs for a controller, especially when a camera prepares the sensory data. Since the captured images have high dimensions, it is important to extract proper task-based features to reduce the dimension of controller input while preserving its performance. In this paper, deep learning is utilized to do this and the appropriate features representing the depth images are acquired. These features besides some expert and memory-based features form a candidate feature set. A feature selection method based on state transition probability is proposed. This method exploits the clustering technique and genetic algorithm. It chooses a subset of candidate feature set to maximize the certainty of transition from the current state of the task environment to the next one. The suggested method named Transition Certainty based Feature Selection (TCFS) is examined for a wheeled robot navigation task in uneven terrains with many obstacles. Wall following and obstacle avoidance tasks are considered as robot missions. Depth images captured by a vision system (Kinect camera) are the only sensory data of the robot controller. The proposed architecture is evaluated in WEBOTS© simulator and on a real robot. The experiments demonstrate that TCFS improves robot performance in comparison to SDL architecture according to defined indexes.