Vehicle Trajectory Tracking With Neural-Tuned MPC Weights
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current trajectory tracking systems for self-driving vehicles rely on predetermined weights and look-up tables for model predictive control, which are inflexible and may not adapt effectively to changing conditions, leading to suboptimal vehicle control.
Innovation Solution
A system that incorporates a trajectory planner, a model predictive controller (MPC) with an internal vehicle model and optimizer, a tuning neural network, and a critic neural network, utilizing reinforcement learning optimization algorithms to dynamically adjust weights and improve control inputs such as steering, throttle, and brake commands.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If predetermined weights and look-up tables are used for model predictive control, then the system structure is simple and easy to implement, but the system lacks flexibility and cannot adapt effectively to changing conditions
Solution Approach 1:
The patent applies the dynamics principle by transitioning from static predetermined weights to dynamic neural network-based weights that adapt in real-time. The neural networks (tuning network and critic network) continuously adjust control parameters based on current system state and performance feedback, enabling the MPC system to adapt to changing conditions while maintaining a structured architecture through the use of trained network models.
Solution Approach 2:
The patent implements parameter changes by modifying the weight parameters of the cost function from fixed predetermined values to dynamically adjusted values generated by neural networks. The tuning neural network and critic neural network continuously optimize the weight matrix Q and other control parameters based on real-time performance evaluation, allowing the system to adapt its control strategy to changing operating conditions.
2Reliability
If reinforcement learning optimization algorithms are used to dynamically adjust weights, then control performance is improved, but computational complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-training the neural networks offline using reinforcement learning algorithms before deployment. The tuning neural network and critic neural network are trained in advance to learn optimal weight adjustments for various operating conditions. During real-time MPC operation, the pre-trained networks provide rapid weight generation based on current state inputs, avoiding the need for complex real-time reinforcement learning computations while maintaining high control performance.
3Manufacturing precision
If iterative weight tuning is performed to minimize the cost function, then trajectory tracking precision is improved, but computation time increases
Solution Approach 1:
The patent substitutes the traditional iterative mechanical optimization process with a neural network-based direct computation approach. Instead of performing iterative numerical optimization to minimize the cost function in real-time, the pre-trained tuning neural network directly generates optimal weights based on current trajectory error and system state. This substitution of iterative optimization with neural network inference dramatically reduces computation time while maintaining high trajectory tracking precision.
Data Source
AI summary
A system for tuning a trajectory tracking controller for a vehicle includes a trajectory planner configured to generate the planned trajectory and to output one or more planned trajectory components representative of the planned trajectory, a model predictive controller including an internal model and an optimizer, and a tuning neural network configured to receive the one or more planned trajectory components and one or more measured trajectory components and to produce weights for a cost function. The internal model is configured to receive a predicted control input from the optimizer and the one or more measured trajectory components and to produce a predicted output. The optimizer utilizes a cost function and is configured to receive the weights for the cost function and a predicted error and to produce the predicted control input, wherein the predicted error is a selected one of the planned trajectory components minus the predicted output.