Vehicle Control Algorithm Training Using Real Traffic Trajectories
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for training algorithms in autonomous vehicles struggle to account for unpredictable behavior of human road users in real traffic situations, as purely virtual training cannot replicate the complexity and variability of real-world interactions.
Innovation Solution
A method involving a self-learning neural network that trains algorithms by comparing actual and virtual trajectories in real traffic environments, using data from environmental and vehicle sensors, and varying traffic situations to expand the spectrum of encountered scenarios, allowing the algorithm to become more robust and adaptive.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the algorithm is trained purely in virtual environments, then training safety and reproducibility are improved, but the algorithm cannot adapt to unpredictable real-world traffic situations
Solution Approach 1:
The patent introduces a human driver as an intermediary between the virtual training environment and real-world deployment. The human driver operates the vehicle in real traffic while the algorithm runs in parallel, creating a bridge that allows the system to gather real-world data without exposing the algorithm to direct real-world risks during the learning phase.
Solution Approach 2:
The system creates a virtual copy of the real-world driving environment by recording sensor data, traffic situations, and human driver decisions. This copied reality is then used to train the algorithm, allowing it to learn from authentic real-world scenarios while maintaining the safety benefits of virtual training.
2Ease of manufacture
If conventional rule-based programming is used for control units, then programming control is simplified, but the system cannot handle complex unpredictable traffic situations
Solution Approach 1:
The patent replaces the mechanical system of conventional rule-based programming with an intelligent system using machine learning algorithms. Instead of manually programming traffic rules and scenarios, the system uses neural networks that automatically learn optimal driving behavior from real-world data, enabling handling of complex unpredictable situations while maintaining ease of deployment.
Solution Approach 2:
The algorithm improves itself automatically through continuous learning from real-world driving data collected during human-operated phases. The system performs self-training and self-optimization without requiring manual reprogramming or intervention, adapting to new traffic patterns and situations autonomously.
3Stability of the object's composition
If the algorithm is frozen after training, then system stability is improved, but continuous improvement through learning is lost
Solution Approach 1:
The patent implements a dynamic training system where the algorithm can transition between stable deployed states and active learning states. During normal operation, the frozen algorithm provides stable control. When specific conditions are met (such as encountering novel situations or scheduled updates), the system activates retraining phases where the algorithm adapts to new data, creating a dynamic balance between stability and adaptability.
Solution Approach 2:
The system employs periodic retraining cycles where the algorithm is temporarily unfrozen to learn from accumulated real-world data, then frozen again for stable operation. This periodic alternation between learning and deployment phases allows continuous improvement while maintaining system stability during critical driving operations.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
Described is a method for training at least one algorithm for a control unit of a motor vehicle, said algorithm being trained by a self-learning neural network, comprising the following steps: a) a computer program product module is provided for the automated or autonomous driving function, b) the trained computer program product module is embedded into the control unit, c) the motor vehicle is driven in a real traffic environment by a human driver, and the journey determines a driven trajectory, d) data from an environmental sensor system and a motor vehicle sensor system are supplied to the control unit, and a virtual trajectory is calculated by the algorithm, e) a metric is derived from a comparison of the driven trajectory and the virtual trajectory, and the data from the environmental sensor system and the motor vehicle sensor system are saved if certain metric criteria are met for a traffic situation, f) information relating to the traffic situation is provided to a traffic simulation, g) the traffic situation is analysed using the traffic simulation, and the data for the traffic situation are varied by means of the traffic simulation, and h) the algorithm is trained by varying the traffic situation.