Closed-Loop Vehicle Decision Control Using Motion Prediction Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The prediction results in the 'prediction-decision control' open-loop solution for intelligent vehicles lack effective guidance for decision control due to limitations in capturing far-distance interactions in space or time, leading to deviations from actual security areas.
Innovation Solution
A prediction-type intelligent vehicle decision control method that establishes an interpretable prediction model for surrounding traffic participants, incorporating an iterative model-driven self-evolution strategy evaluation and strategy promotion process, using graph neural networks and attention mechanisms to update and embed the prediction model into a decision control system.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a prediction-decision control open-loop solution is used, then the decision control process is simplified, but the prediction results lack effective guidance for decision control
Solution Approach 1:
The patent implements a closed-loop feedback mechanism where prediction results are fed back into the decision control process. The decision control module receives prediction results from the prediction module and uses them to generate control decisions, which are then executed and fed back to update the prediction model, creating a continuous improvement cycle that enhances both ease of operation and reliability
Solution Approach 2:
The patent merges the prediction module and decision control module into an integrated system where both functions operate simultaneously with shared data flows. The prediction module and decision control module are combined in a unified architecture that allows real-time interaction between prediction results and control decisions, eliminating the separation inherent in open-loop solutions
2Measurement precision
If rasterizing encoding of map information is used, then spatial information is captured, but far-distance interaction in space or time cannot be captured
Solution Approach 1:
The patent transitions from traditional rasterizing encoding to a graph-based representation that adds a topological dimension to spatial information. By representing map information as a graph structure with nodes and edges, the system can capture far-distance interactions through graph neural network message passing that propagates information across the entire graph, overcoming the limited receptive field of local convolutional operations
Solution Approach 2:
The patent replaces the mechanical rasterizing encoding mechanism with a graph neural network-based information propagation mechanism. Instead of using fixed-grid rasterization with limited local receptive fields, the system uses graph-based message passing that can dynamically propagate information across arbitrary distances in the graph structure, enabling capture of far-distance spatial and temporal interactions
3Measurement precision
If memory mechanism is used, then temporal information is captured, but prediction results may deviate from actual security area
Solution Approach 1:
The patent uses feedback to continuously refine predictions by comparing predicted trajectories with actual observed trajectories. The loss function calculates deviations between predicted and actual positions, and this error information is fed back to update the model parameters through backpropagation, improving prediction accuracy over time and ensuring predictions remain within actual security areas
Solution Approach 2:
The patent implements dynamic adaptation of the prediction model through continuous training and updating. The model parameters are dynamically adjusted based on real-time feedback from the environment, allowing the system to adapt to changing traffic patterns and improve prediction accuracy. The training process uses dynamic gradient descent to optimize parameters, making the prediction system flexible and responsive rather than static
Data Source
AI summary
A prediction-type intelligent vehicle decision control method and apparatus, a vehicle, and a storage medium are provided. The method includes: establishing a prediction model for surrounding traffic participants based on map information and historical tracks of the traffic participants, and performing parameter initialization on the prediction model by using a labeled data set to generate an initial surrounding vehicle motion prediction model; cyclically updating the initial surrounding vehicle motion prediction model according to continuous interaction data with an environment by taking a driving target of an intelligent vehicle as an optimization object, so as to generate a final surrounding vehicle motion prediction model; and embedding the final surrounding vehicle motion prediction model into a decision control system of the intelligent vehicle, the decision control system generating a decision control instruction according to a predicted surrounding vehicle motion and controlling the intelligent vehicle to execute the decision control instruction.