Deep reinforcement learning training acceleration method for collision avoidance of multiple unmanned aerial vehicles

Abstract

The invention discloses a deep reinforcement learning training acceleration method for collision avoidance of multiple unmanned aerial vehicles. The deep reinforcement learning training acceleration method comprises the steps of S1, performing formalized modeling on a completely distributed unmanned aerial vehicle cluster obstacle avoidance problem based on a Markov decision process; S2, constructing a deep neural network, and constructing a mapping and network updating method of observation input-action output; and S3, fusing human experience to accelerate training. The method has the advantages that the principle is simple, a training intelligence degree is high, the deep reinforcement learning training process can be accelerated and so on.

Description

technical field [0001] The invention mainly relates to the technical field of unmanned aerial vehicles, in particular to a deep reinforcement learning training acceleration method for collision avoidance of multiple unmanned aerial vehicles. Background technique [0002] With the gradual expansion of the application field of UAVs, people's demand for UAVs to perform tasks autonomously is getting higher and higher. Autonomous positioning, environment perception, path planning and collision avoidance are key technologies for UAVs to perform tasks autonomously. Compared with single UAVs, multi-UAVs can carry more mission loads, have a larger detection range, and perform various tasks. [0003] Deep reinforcement learning not only has the ability of deep learning to understand complex high-dimensional data, but also has the general learning ability of reinforcement learning to learn by itself through trial and error mechanisms. However, deep reinforcement learning mostly faces...

AI Technical Summary

Problems solved by technology

The methods based on transfer learning are mostly used in visual information perception tasks, such as target recognition, etc., but for sensors such as lidar, the trained network model and parameters cannot be directly transferred and applied

[0006] (2) The degree of automation is not high

Staged training breaks down a certain task into multiple stages of tasks, and trains them sequentially, which is more cumbersome

Moreover, in staged training, subsequent stages of training may lead to policy forgetting in the pre-training stage

However, most of the existing human-guided training requires people to participate in the training process as teachers, which has a low degree of automation and consumes a lot of time and energy for developers.

Images