End-to-End Neural Network for Autonomous Vehicle Trajectory Control

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional autonomous driving systems using modular approaches struggle to transition smoothly between driving commands and often fail to learn the dynamics of driving due to strong data correlation between commands and output trajectories, limiting their ability to perform human-like maneuvers and navigate unpredictably.

Innovation Solution

A single end-to-end artificial neural network is trained using a combination of definitive and random mode commands, where definitive commands are input along with neutral commands to induce 'confusion', allowing the network to learn the overall dynamics of driving and maintain control over trajectory choices, enabling smooth human-like steering behavior and obstacle avoidance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a single end-to-end neural network is used to control vehicle driving, then the system can learn overall driving dynamics and produce smooth trajectories, but the network struggles to transition smoothly between driving commands due to strong data correlation between commands and output trajectories

Engineering Contradiction:

Improveability to learn driving dynamicsVSAvoidsmooth transition between commands

Core Design Contradiction:

Adaptability or versatilityVSEase of manufacture

Solution Approach 1:

The training process is segmented into two distinct modes: definitive mode training and random mode training. Definitive mode training teaches the network specific driving commands (straight, left, right), while random mode training induces confusion by randomly selecting commands to improve smooth transitions. This segmentation resolves the contradiction by addressing different aspects of driving behavior in separate training phases.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The training approach dynamically adjusts between two operational modes. The network learns to adapt its behavior based on the training mode: in definitive mode it learns precise command execution, while in random mode it learns to handle transitions smoothly. This dynamic training strategy enables the network to master both specific command execution and smooth transitions between commands.

Inventive Principle:

Principle #15Dynamics

2Measurement precision

If definitive commands are trained with strong correlation between commands and trajectories, then the network learns specific driving maneuvers, but it fails to learn overall driving dynamics and produce human-like maneuvers

Engineering Contradiction:

Improvecommand execution accuracyVSAvoidhuman-like driving behavior

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

Random mode training acts as a preliminary anti-action to the strong correlation problem in definitive mode training. By intentionally introducing confusion and weakening command-trajectory correlations during random mode training, the network learns to generalize beyond specific commanded trajectories, enabling it to produce more human-like and adaptable driving behavior while maintaining command execution accuracy from definitive mode training.

Inventive Principle:

Principle #9Preliminary anti-action

Data Source

PatentEP3717981B1Methods for training and controlling an artificial neural network with discrete vehicle driving commands

Publication Date: 2023.07.12 VOLKSWAGEN AG

Figure 1
Figure 2
Figure 3

AI summary

Systems, devices, and methodologies are provided for training and controlling a neural network. The neural network is trained using definitive and random training modes to train neurons in a monolithic network. The neural network output is used to control an autonomous or semi-autonomous vehicle.