Autonomous Driving Policy Training for Extreme-Scene Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous driving systems lack a comprehensive and safe driving control policy, especially in extreme scenes, due to insufficient training data and the danger and costliness of independent explorations in real-world environments.
Innovation Solution
A training system comprising a simulator module to simulate static and dynamic factors, a policy search module using machine learning to optimize driving control policies, and a policy transfer module to adapt policies for unmanned vehicles, which includes constructing a simulator, detecting and predicting dynamic factors, designing an objective function, and updating policies through evolutionary or reinforcement learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If supervised learning with large amount of driving data is used to train the model, then the model outputs can be similar to human driving habits, but the model still cannot cover all driving scenes especially extreme scenes due to lack of extreme scene data in collected data
Solution Approach 1:
The patent creates a virtual copy of the real driving environment through a simulator that replicates road scenes, vehicle dynamics, and especially extreme scenarios that are difficult to capture in real data. This virtual environment allows the model to be trained on diverse scenes including rare extreme cases without needing to collect equivalent real-world data.
Solution Approach 2:
The patent performs preliminary training in a simulated environment before deploying to real-world applications. The simulator pre-generates training data including extreme scenes, and the model undergoes preliminary learning in this controlled virtual setting, which then transfers to improved performance in real autonomous driving scenarios.
2Extent of automation
If reinforcement learning with interactive trial-and-error is used to improve decision-making capacity, then the intelligent agents can learn optimal control policy, but unmanned vehicles are required to perform large quantity of independent explorations in physical world which is extremely dangerous and costly
Solution Approach 1:
The patent introduces a simulator as an intermediary between the reinforcement learning agent and the real physical world. The agent performs all trial-and-error interactions within this virtual intermediary environment, learning optimal control policies without direct exposure to real-world dangers. The simulator mediates the learning process, allowing safe exploration of extreme scenarios.
Solution Approach 2:
The patent creates a virtual copy of the physical driving environment where reinforcement learning can occur safely. This simulated副本 reproduces the dynamics and challenges of real driving, enabling the agent to learn through trial-and-error without risking physical safety or incurring real-world costs.
3Ease of manufacture
If manual rule schemes or real-time planning schemes are used for driving control, then the system can operate without extensive training data, but the control policy is not intelligent and has serious defects in realizing safe driving
Solution Approach 1:
The patent replaces manual rule-based mechanical control systems with an intelligent machine learning model. Instead of relying on pre-programmed rules and real-time planning algorithms, the system uses a neural network trained in simulation to automatically learn and execute safe driving behaviors, achieving both intelligence and safety.
Solution Approach 2:
The patent performs preliminary learning and policy development in a simulated environment before deployment. The intelligent control policy is pre-trained extensively in virtual scenarios including extreme cases, so that when deployed in the real vehicle, it already possesses sophisticated safety capabilities without requiring complex real-time rule processing.
Data Source
AI summary
The invention discloses a training system for autonomous driving control policy, which comprises a simulator construction module based on machine learning, a driving control policy search module based on confrontation learning, and a driving control policy model transfer module.