Unmanned Vehicle Path Planning With Deep RL Map Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current global path planning methods for unmanned vehicles face challenges such as high computational costs, reliance on manual marking, and poor model performance, especially in dynamically changing environments, due to difficulties in modeling and training neural network models.
Innovation Solution
A global path planning method and device utilizing a reinforcement learning approach with a deep reinforcement learning neural network, where an object model describes the sequential decision-making process incorporating the state of the unmanned vehicle and environmental map pictures, allowing for the generation of motion paths based on evaluation indices, and utilizing a multi-GPU framework for improved training efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep reinforcement learning neural network is used for path planning, then adaptability to dynamic environments is improved, but training time is excessively long
Solution Approach 1:
The patent pre-processes map pictures into feature maps before training, extracting key environmental features in advance. This preliminary action reduces the complexity of the input data during training, allowing the neural network to learn more efficiently and reach stable performance faster, thereby reducing training time while maintaining adaptability to dynamic environments
Solution Approach 2:
The patent transforms the raw map picture into a feature map by changing the representation parameters from pixel values to extracted environmental features. This parameter transformation simplifies the input data structure, enabling faster convergence during training while preserving the essential information needed for adaptability to dynamic changes in the environment
2Measurement precision
If deep reinforcement learning neural network is used for path planning, then path planning accuracy is improved, but computational cost is high
Solution Approach 1:
The patent extracts essential environmental features from complete map pictures, separating only the necessary information (obstacles, corridors, key landmarks) from unnecessary details. This extraction reduces the computational load on the neural network while preserving the critical information needed for accurate path planning, thereby lowering computational cost without sacrificing accuracy
3Manufacturing precision
If manual marking is used for trajectory preparation, then training data quality is improved, but labor intensity is high
Solution Approach 1:
The system automatically generates training data by having the trained neural network perform path planning on pre-processed feature maps. The network learns through self-play and automatic evaluation, eliminating the need for manual trajectory marking. This self-service approach maintains high training data quality through automated feedback mechanisms while completely removing manual labor
Solution Approach 2:
Instead of manually creating training trajectories, the system uses the neural network to generate and evaluate paths automatically, copying the decision-making process from the trained model back into the training data generation. This creates a closed-loop system where the model improves itself through automated experimentation, maintaining data quality without human intervention
Data Source
AI summary
A global path planning method and device for an unmanned vehicle are disclosed. The method comprises: establishing an object model through a reinforcement learning method, wherein the object model includes: a state of the unmanned vehicle, an environmental state described by a map picture, and an evaluation index of a path planning result; building a deep reinforcement learning neural network based on the object model established, to obtain a stable neural network model; inputting the map picture of the environment state and the state of the unmanned vehicle into the deep reinforcement learning neural network after trained, and generating a motion path of the unmanned vehicle. According to the present disclosure, the environment information in the scene is marked through the map picture, and the map features are extracted through the deep neural network, thereby simplifying the modeling process of the map scene.