TDPP-Net: Achieving three-dimensional path planning via a deep neural network architecture

BND*-DDQN: Learn to Steer Autonomously Through Deep Reinforcement Learning

“…According to the target Q-values and their current values estimated by the online network, our loss function is then defined as: (13) where α 1 , α 2 , and α 3 are three scaling factors. The first two terms in Equation (13) are aimed at minimizing the differences between the predicted Q-values and their corresponding target values.…”

“…According to the target Q-values and their current values estimated by the online network, our loss function is then defined as: (13) where α 1 , α 2 , and α 3 are three scaling factors. The first two terms in Equation (13) are aimed at minimizing the differences between the predicted Q-values and their corresponding target values. And because the angular and linear velocities are performed concurrently during the interaction, the last term is designed to minimize the difference between the Q-value estimates corresponding to the two commands.…”

“…In recent years, approaches developed based on various deep learning techniques have been introduced to address the bottlenecks of conventional methods and explore solutions to the aforementioned issues [10]- [13]. As an important category, deep reinforcement learning (DRL) based methods have gained rapidly growing popularity since they demonstrate promising performance while impose no requirement on labeled data.…”

Waypoint Planning Networks

“…Learning-based planning algorithms have increasingly become more common [23], [24], [25], [26]. In most learningbased path planning algorithms, imitation learning [27] plays a key role [28]. Neural networks have been used to improve the classic algorithms, for instance by adaptively sampling a particular region of a configuration space in sampling-based algorithms [29].…”

“…Reinforcement Learning (RL) approaches such as value iteration networks (VINs) [13], [36], learning-fromdemonstration (LfD) [37], guided policy search (GPS) [36], and universal planning networks (UPN) [38] have also been used for path planning. Wu et al present three-dimensional path planning network (TDPP-Net) [28], which is an end-toend network that predicts 3D actions via 2D CNNs. TDPP-Net learns a policy via supervised imitation learning from the Dijkstra's algorithm.…”