Curriculum Sequence Reinforcement Learning for Autonomous Vehicle Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in navigating complex urban environments due to the abundance of external factors and unpredictable changes, making it difficult to design reliable planning algorithms for safe decision-making, especially at intersections, as prior distance-based algorithms require laborious tuning and are not easily adaptable to various scenarios.
Innovation Solution
The implementation of an automatically generated curriculum sequence based reinforcement learning method, which trains a vehicle control system to perform tasks in order of difficulty, using a processor to determine tasks and generate an action policy to maximize rewards, allowing the vehicle to learn and adapt to environmental changes without losing crucial details in computation complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If distance-based algorithms are used for intersection traversing, then navigation decisions can be made, but extensive tuning and rule design are required which increases system complexity
Solution Approach 1:
The reinforcement learning agent autonomously learns optimal navigation policies through self-interaction with the environment, eliminating the need for manual tuning of distance-based parameters and rule design. The system serves itself by automatically adapting to various intersection scenarios through trial-and-error learning.
Solution Approach 2:
The patent transitions from fixed distance-based parameters requiring manual tuning to learned policies that dynamically adjust behavior based on environmental feedback. The reinforcement learning framework changes the approach from static parameter setting to adaptive parameter optimization through training.
2Reliability
If all available perception data is used for decision making, then situation awareness is improved, but computation complexity increases
Solution Approach 1:
The patent extracts only the most relevant features from the full perception data through the reinforcement learning training process. The agent learns to focus on critical information for navigation decisions, effectively filtering out redundant data and reducing computation complexity while maintaining situation awareness.
Solution Approach 2:
The complex perception data is segmented into meaningful features that the reinforcement learning agent can process efficiently. The training process divides the overwhelming amount of sensor data into relevant components for decision-making, reducing computational burden.
Data Source
AI summary
The present application generally relates to a method and apparatus for generating an action policy for controlling an autonomous vehicle. In particular, the system performs a deep learning algorithm in order to determine the action policy and an automatically generated curriculum system to determine a number of increasingly difficult tasks in order to refine the action policy.