Vehicle AI Training Data Using Predicted Trajectory Image Sequences

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current AI training methods for autonomous robots and vehicles rely on simplified simulations, which limit the realism and effectiveness of training data, leading to suboptimal performance in real-world scenarios due to high computing efforts and reduced environmental representation.

Innovation Solution

A method that generates realistic training data records by recording image sequences of real environments, determining trajectories, and predicting future movements to create high-quality training datasets for AI modules, using a pseudo or offline simulator that reduces computing effort while maintaining realism.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a realistic simulator is used to train the AI module with sufficient environmental representation, then the training quality and AI performance are improved, but the computing effort increases significantly

Engineering Contradiction:
Improvetraining qualityVSAvoidcomputing effort
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent segments the training process into two distinct phases: (1) an offline phase where a realistic simulator pre-generates training data sets with full environmental representation, and (2) an online phase where the AI module is trained using this pre-generated data. This segmentation allows the computationally intensive simulation work to be performed once offline, while the actual training online requires minimal computing resources.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies preliminary action by pre-generating comprehensive training data sets using a realistic simulator before the actual AI training process. The simulator pre-computes environmental representations, object behaviors, and training scenarios in advance, storing them as ready-to-use training data sets that can be directly fed to the AI module during training without requiring real-time simulation computing power.

Inventive Principle:
Principle #10Preliminary action

2Use of energy by moving object

If the environment is reduced or simplified to a model for simulation, then the computing effort is reduced, but the training success of the AI module deteriorates

Engineering Contradiction:
Improvecomputing effortVSAvoidtraining success
Core Design Contradiction:
Use of energy by moving objectVSReliability

Solution Approach 1:

The patent creates a copy of the real environment through a realistic simulator that generates virtual training data sets. This copy includes accurate representations of roads, static objects, dynamic objects, and their interactions. The simulator copies essential environmental features and behaviors into virtual scenarios, providing the AI module with training data that faithfully represents real-world conditions without requiring physical presence in those environments.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent performs preliminary action by using the realistic simulator to pre-generate high-quality training data sets before AI training. This preliminary simulation phase creates comprehensive environmental models including diverse road types, weather conditions, traffic patterns, and object behaviors, ensuring the training data adequately represents the complexity of real environments while avoiding the need for simplified models during actual training.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If a long time segment is predicted in the future image sequence, then the realism and completeness of training scenarios are improved, but the computing effort increases

Engineering Contradiction:
Improvetraining scenario realismVSAvoidcomputing effort
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent applies partial action by predicting only the necessary portion of future image sequences required for effective AI training. Rather than generating excessively long future predictions, the system determines an optimal prediction horizon that provides sufficient training scenarios for the AI module to learn safe and effective behavior, avoiding unnecessary computation beyond what is needed for training adequacy.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3765927B1Method for generating a training data record for training an artificial intelligence module for a control device of a vehicle
Publication Date: 2023.01.04 ROBERT BOSCH GMBH
  • EP3765927B1 patent drawingFigure 1
  • EP3765927B1 patent drawingFigure 2

AI summary

A method is provided for generating a training data record for training an artificial intelligence (AI) module (1). In the method, an image sequence (5), in which surroundings (6) of a robot are detected, is first provided. At least one trajectory (14a-14e) that can be arranged in the detected surroundings (6) of the robot is then determined. In the method, at least one future image series is also generated, which image series extends to a time segment (t0+n) in the future in relation to a sequence end time (t0), and images are predicted, based on the at least one particular trajectory (14a-14e), for the case in which the particular trajectory (14a-14e) would be followed during the future time period segment (t0+n). At least one partial section of the trajectory (14a-14e) determined contained in the image series generated is evaluated as positive if a movement predicted by following the trajectory (14a-14e) corresponds to a valid movement situation, or as negative if the movement predicted by following the trajectory (14a-14e) corresponds to an invalid movement situation; and the future image series generated is combined with said allocated evaluation of the trajectory (14a-14e) to generate a training data record (2) for the AI module (1). As a result, the AI module (1) can be trained by a driving simulator that is based on the recorded and thus realistic image sequence (5) in combination with a prediction that can be obtained with relatively low computing effort.