Automatic driving method based on conditional imitation learning and reinforcement learning

Abstract

The invention provides an automatic driving method based on conditional imitation learning and reinforcement learning, and the method comprises the steps: S1, collecting driving data, and marking a corresponding decision action as an expert decision; s2, according to the driving data, calculating to obtain a course angle, and obtaining a feature vector and a feature map; s3, constructing a training data set based on the feature vector, the feature map, the course angle and the expert decision, and training a model by using the training data set; and S4, applying the trained model to the vehicle to realize the automatic driving decision of the vehicle. According to the method, by combining conditional imitation learning and reinforcement learning, the dependence of the model on a training data set is reduced, and meanwhile, the problem of low reinforcement learning exploration efficiency caused by random initialization is solved; real vehicle data are collected and applied to model training, so that the trained model is closer to a real driving scene.

Description

technical field [0001] The invention relates to the field of automatic driving, in particular to an automatic driving method based on conditional imitation learning and reinforcement learning. Background technique [0002] The application of autonomous driving can significantly reduce traffic accidents, alleviate traffic congestion, improve traffic efficiency, and save energy consumption, so it has been highly valued. Governments, enterprises, and research institutions have invested huge amounts of money and manpower in the research of autonomous driving, aiming to realize its commercial application as soon as possible. The traditional rule-based (Rule-Based) automatic driving solution subdivides the entire system into modules such as perception, positioning, planning, and control. Through the combined use of these modules, the automatic driving of the vehicle is realized. However, this scheme has a complex structure, high cost, and there are multiple intermediate links, an...

AI Technical Summary

Problems solved by technology

[0008] However, the methods proposed in the above papers and patents have the following deficiencies: simply using RGB images as the main source of perception has poor interpretability and robustness. Very unsafe; the rough PP angle calculated by image point selection is too rough, and a better control amount of real vehicle data can be selected in actual landing; the imitation learning framework has a huge convolutional layer and insufficient real-time performance

However, this invention does not reduce the model's dependence on the training data set by combining CIL and RL, nor does it solve the problem of low efficiency of reinforcement learning exploration due to random initialization.

Images