Autonomous Vehicle MPC Planning with Learned Constraint Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle decision-and-planning modules face challenges in ensuring optimal driving performance in diverse and complex driving situations due to difficulties in handling non-convex constraints, leading to local optimal solution convergence and real-time performance issues.
Innovation Solution
The method employs a real-time model predictive control-based planning system that uses identification and classification learning of constraints through a deep learning network to simplify non-convex constraints into convex constraints, enabling robust and efficient trajectory planning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If optimization-based planning system is used to find optimal trajectory achieving constraints, then driving performance and stability are improved, but computational amount increases and real-time application becomes difficult
Solution Approach 1:
The planning system is divided into two segments: a deep learning-based decision module that handles high-level decision-making and constraint classification, and an optimization-based planning module that handles trajectory generation. This segmentation allows the computationally intensive optimization to operate on simplified inputs from the fast neural network, achieving both optimal trajectories and real-time performance.
Solution Approach 2:
The deep learning network performs preliminary classification of constraints and prediction of decision values before the optimization process. By pre-processing the problem space and identifying relevant constraints in advance, the optimization algorithm receives refined inputs that reduce its computational burden while maintaining solution quality.
2Productivity
If rule-based planning system is used for intuitive and efficient planning, then computational amount is reduced, but ability to handle all situations and controllability deteriorates
Solution Approach 1:
The deep learning network serves as an intermediary between the rule-based constraint definitions and the optimization planner. It learns to generalize constraint patterns from training data, enabling the system to handle diverse and complex situations that would be difficult to encode with explicit rules, while still maintaining computational efficiency through the neural network's parallel processing capability.
3Productivity
If learning-based planning system is used for complex driving environments, then real-time performance is improved through parallel processing, but ability to learn various situations and stability deteriorates
Solution Approach 1:
The system merges the strengths of deep learning and optimization-based planning into a unified hybrid architecture. The deep learning component provides fast real-time inference through parallel processing, while the optimization component ensures stable and reliable trajectory generation. This combination leverages the speed of neural networks and the reliability of optimization algorithms to achieve both real-time performance and stability.
Data Source
AI summary
Disclosed is a method and system for real-time model predictive control-based planning of an autonomous vehicle based on identification and classification learning of constraints. A real-time model predictive control-based planning method may include generating a class prediction decision value that determines an upper bound and a lower bound of constraints in an optimal control problem through an observation value for a driving environment; and receiving the class prediction decision value and the observation value from the driving environment and generating a main trajectory.