Open AccessPolicy-based monocular vision autonomous quadrotor obstacle avoidance method
Author(s) -
Baojin Zheng,
Xiao Guo,
Jiajun Ou
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/2083/3/032025
Subject(s) - obstacle avoidance , robustness (evolution) , obstacle , computer science , generalization , reinforcement learning , monocular , collision avoidance , artificial intelligence , control theory (sociology) , robot , control (management) , mathematics , mobile robot , computer security , mathematical analysis , biochemistry , chemistry , political science , law , gene , collision
Aiming at the obstacle avoidance control problem of small quadrotor, a method of quadrotor obstacle avoidance based on reinforcement learning is proposed. The proposed method can make training converge quickly and has good environmental robustness. The proposed methods include: (1) a framework adopts perception module and decision module to improve the generalization ability of the obstacle avoidance model; (2) An Actor-Critic framework-based Proximal Policy Optimization (PPO) algorithm to provide quadrotor with policy-based decision-making capabilities; The experimental simulation results show that the strategy-based framework converges quickly and has a high success rate, the training time is much lower than that of the value-based framework. The monocular vision observation ability is limited, which leads to deviations between local observation and global state, So LSTM layer is usually added to increase model performance. Policy -based decision can have a good obstacle avoidance effect without adding the LSTM layer, and have good generalization ability after short relearning after changing.