Towards Monocular Vision based Obstacle Avoidance through Deep Reinforcement Learning

A Vision Based Deep Reinforcement Learning Algorithm for UAV Obstacle Avoidance

“…Generally, these states are represented as depth prediction images and the size of the reward influences the strength of the signal for which the UAV will alter its behavior policy. For the problem of obstacle avoidance, [1] demonstrated that deep Q-learning as introduced by [15], is effective for robotic navigation in 2-Dimensions. Alternative approaches have been tried including work by Khan et al [22] provided a model-based approach that estimates the probability that the UAV would have a collision within an unknown environment, allowing for specific actions depending on the certainty of the prediction.…”

“…One of the major challenges with autonomous motion planning is ensuring that an agent can efficiently explore a space while avoiding collision with objects in complex and dynamic environments. To resolve this, recent research has turned to applying deep reinforcement learning techniques to robotics and UAVs [1], [2].…”

“…This iterative feedback loop of receiving signals through a sensor to reflect the agents current state allows it to train an optimal policy. Part of the research in this area has focused on employing a single sensor such as a camera [1], [3] to control an agent, while others have focused on combining the signals from multiple sensors [4]. Object detection algorithms have shown tremendous success in the past few years for both detection and segmentation in computer vision.…”

“…Particularly, when an agent navigates through an environment and learns from the state transitions, it may rarely receive any significant reward or signal from the environment in which to learn from, resulting in a significantly slower learning rate [8]. The Dueling Double Deep Q Network (D3QN) algorithm for obstacle avoidance introduced by [1] uses the -greedy or Deep Q-Network (DQN) approach introduced by Mnih et al [9] for choosing between exploration or exploitation. That is, either exploring through a stochastic process (exploration) or exploiting the behavioral policy learned so far in order to chose the next action.…”

Deep Reinforcement Learning for Mapless Navigation of Unmanned Aerial Vehicles

“…More recently, the learning was applied in tasks in robotics as well, where it was initially used to handle tasks in stable and observable environments [3]. For mobile robotics, however, the complexity increases significantly given the interactions with barriers in the physical workplace [4,5]. In this context, Deep-RL ended up simplifying the problem by discretizing it [6].…”