Unmanned vehicle reinforcement learning training environment construction method and training system thereof

Abstract

The invention discloses an unmanned vehicle reinforcement learning training environment construction method and a training system thereof, and belongs to the field of robot navigation and the field ofrobot simulation platforms. The method comprises the following steps: constructing a real scene and simulation scene data set; enhancing a data set; training an image domain conversion algorithm andstoring a model; and establishing an API interface of the simulation environment model and the reinforcement learning algorithm. During training in a simulation environment, a camera on the unmanned vehicle model collects an observed simulation environment image, the simulation environment image is converted into a simulated real scene image through an image domain conversion network, the simulated real scene image serves as a state to be input into a reinforcement learning network, and an action instruction is output through decision making and issued to the unmanned vehicle model of a simulation end. In practical application, the unmanned vehicle camera collects real scene pictures in reality, and the simulated real scene pictures input by the reinforcement learning algorithm in trainingare very similar to the real scene pictures, so that the trained algorithm can be directly migrated or migrated to a real scene after fine adjustment.

Description

technical field [0001] The invention relates to the field of robot navigation and robot simulation platform, in particular to a method for constructing an unmanned vehicle intensive learning training environment and a training system thereof. Background technique [0002] Robot navigation is to get it from an initial position to a target position without colliding with obstacles during the process. Traditional robot navigation is based on maps. If the environment model (map) is known in advance, it is a global path planning problem. This method has low requirements on the computing power of the robot system and can find the optimal solution. If the environment model is unknown in advance or only locally known, the robot needs to perceive the surrounding environment through sensors such as lidar or camera fixed on it during the movement, and perform real-time modeling and correction of the surrounding environment. This method is called Local path planning problem. Local pa...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the defect in the prior art that it is difficult to directly migrate the reinforcement learning strategy trained in the simulation environment to the real environment, the present invention provides a method for constructing an unmanned vehicle reinforcement learning training environment and its training system, using the image domain conversion algorithm, Convert the pictures in the training environment into pictures in the simulated real scene, and input them into the reinforcement learning algorithm as the state, and then obtain the decision-making of each step of the unmanned vehicle

Images