A virtual-to-real transfer method, control method and related system for an autonomous vehicle decision-making model

By constructing a pose dataset and a neural radiation field network, a mapping relationship between virtual and real objects is established, and a semantic graph of the real task environment is constructed. This solves the problem of the lack of correlation between virtual and real environment inputs and realizes seamless virtual-real migration of the autonomous vehicle decision-making algorithm model.

CN118095409BActive Publication Date: 2026-06-30SHANGHAI UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
SHANGHAI UNIV
Filing Date
2024-03-08
Publication Date
2026-06-30

AI Technical Summary

Technical Problem

In existing virtual-real transfer methods, the inputs of the virtual environment and the real environment are not correlated, which means that valuable data from the real environment is needed to fine-tune the decision model when training it. Furthermore, it is difficult to solve the problem of cross-scene transfer of the same task, resulting in a sharp decline in policy performance.

Method used

By constructing a pose dataset, using a neural radiation field network to learn the geometric features of virtual objects, establishing a mapping relationship between the real task environment and virtual objects, constructing a semantic graph of the real task environment, and using it to train an autonomous vehicle decision-making algorithm model, thus achieving seamless virtual-real transfer of the decision-making algorithm model.

Benefits of technology

There is no need to train the autonomous vehicle in a real-world environment. The pre-trained decision-making algorithm model can be directly deployed to the autonomous vehicle, solving the problem of cross-scene migration for the same task and achieving seamless migration of the decision-making algorithm model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118095409B_ABST
    Figure CN118095409B_ABST
Patent Text Reader

Abstract

This invention discloses a method, control method, and related system for virtual-real transfer of an autonomous vehicle decision-making model, relating to the fields of deep learning and intelligent unmanned systems. The method includes: learning the features of virtual objects using a neural radiation field network based on a virtual object pose dataset to obtain a neural network model corresponding to each type of virtual object; establishing a mapping relationship between each real object and each type of virtual object in the real task environment, and determining the neural network model corresponding to each real object accordingly; constructing a semantic graph of the real task environment; training a decision-making algorithm model based on the semantic graph; and deploying the trained decision-making algorithm model onto the controller of the autonomous vehicle. This invention achieves seamless virtual-real transfer of the decision-making algorithm model by associating real objects with virtual objects in the task environment, thereby constructing a consistent virtual-real task environment semantic graph, which is then used to train the autonomous vehicle decision-making algorithm model.
Need to check novelty before this filing date? Find Prior Art