Autonomous Driving Algorithm Training With Real-Object Traffic Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training algorithms for autonomous vehicles are inadequate in preparing them for real-world traffic scenarios due to the inability to account for characteristics and wear in detection devices, which cannot be modeled in purely virtual training.
Innovation Solution
A method involving self-learning neural networks trained on a test site with real objects moved by remote control, where the algorithm is trained in a simulation environment connected to real objects, allowing for realistic training without endangering human life and enabling continuous improvement during everyday operation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data is recorded directly in the motor vehicle during actual operation, then the training data reflects real operating conditions and improves algorithm accuracy, but the recording process increases device complexity and requires additional storage and processing resources in the vehicle
Solution Approach 1:
A server acts as an intermediary between the motor vehicle and the algorithm training process. The server receives training data from the vehicle, performs the algorithm training, and provides trained algorithms back to the vehicle. This mediator approach allows the vehicle to collect real operating data without needing complex local training capabilities, resolving the contradiction between data quality and device complexity.
Solution Approach 2:
The algorithm training function is extracted from the motor vehicle and relocated to an external server. The vehicle only needs to perform data collection and transmission, while the computationally intensive training process occurs externally. This extraction reduces the vehicle's device complexity while maintaining access to high-quality real-world training data.
2Productivity
If training data is collected from simulation environments, then the training process is faster and requires fewer resources, but the training data lacks the complexity and variability of real-world operating conditions
Solution Approach 1:
The system merges simulation-generated training data with real-world recorded training data. Simulation data provides large volumes of diverse scenarios including rare events, while real-world data ensures authenticity and accuracy. This combination approach achieves both high productivity from simulation and high measurement precision from real data.
Solution Approach 2:
Simulation environments are used to perform preliminary training and generate initial training datasets before real-world data collection. This preliminary action allows the algorithm to learn basic patterns quickly from simulated data, then be refined using real-world data, combining the speed of simulation with the accuracy of real measurements.
3Quantity of substance
If extensive training data is stored in the motor vehicle for later processing, then more comprehensive training scenarios are available, but the storage requirements and data management complexity increase significantly
Solution Approach 1:
The server acts as an intermediary that centralizes data management functions. Instead of the vehicle storing and managing large datasets locally, the vehicle transmits data to the server which handles storage, organization, and retrieval. This reduces vehicle complexity while enabling access to extensive training data volumes through the intermediary's centralized management system.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A method for training at least one algorithm for a control unit of a motor vehicle is described, wherein the control unit is provided for implementing an autonomous driving function with intervention in assemblies of the motor vehicle on the basis of input data by using the at least one algorithm, wherein the algorithm is trained by means of a self-learning neural network, the method comprising the following steps: a) providing a computer program product module for the autonomous driving function, wherein the computer program product module contains the algorithm to be trained and the self-learning neural network; b) embedding the trained computer program product module into the control unit of the motor vehicle, wherein c) a simulation of critical traffic situations relevant to the autonomous driving function is carried out at a test site at which real objects are moved under remote control, and d) the algorithm is trained by self-learning in the motor vehicle.