Neural Network Semantic Map Correction via Simulation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing technologies rely heavily on manual operations and mapping experts to correct and update semantic grid maps, which are prone to errors due to sensor inaccuracies and ambiguities, leading to increased costs and efforts in maintaining accurate maps.

Innovation Solution

A system and method that utilize a neural network to learn how to build and update large-scale consistent semantic grid maps by simulating a robot's movement in a generated environment, comparing actual and predicted geographical information, and correcting flaws without manual intervention.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual operations are used to correct and update semantic grid maps, then map accuracy can be maintained, but costs and efforts increase significantly

Engineering Contradiction:
Improvemap accuracyVSAvoidmaintenance costs and efforts
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The system enables self-service by training the neural network to automatically detect and correct errors in semantic grid maps through simulated robot experiences. The network learns to identify inconsistencies between sensor data and map representations, and autonomously generates corrections without human intervention, thereby maintaining map accuracy while eliminating manual labor costs.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent replaces the mechanical manual correction process with an automated neural network system. Instead of mapping experts manually reviewing and correcting map errors, the system uses a trained neural network that processes sensor data and automatically generates map corrections, substituting human mechanical operations with an automated computational system.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Stability of the object's composition

If manual expert correction is applied to semantic grid maps, then map consistency is improved, but time consumption increases

Engineering Contradiction:
Improvemap consistencyVSAvoidtime consumption for corrections
Core Design Contradiction:
Stability of the object's compositionVSLoss of time

Solution Approach 1:

The system applies preliminary action by pre-training the neural network on extensive simulated robot experiences before deployment. During operation, the already-trained network can immediately detect and correct map inconsistencies without requiring manual expert intervention, thereby maintaining map consistency while minimizing time consumption compared to waiting for manual corrections.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system implements feedback mechanisms where the neural network continuously compares sensor observations with map representations, identifies inconsistencies, and generates corrections. This closed-loop feedback process automatically maintains map consistency by detecting and correcting errors as they occur, eliminating the time delay associated with manual expert review cycles.

Inventive Principle:
Principle #23Feedback

3Ease of manufacture

If semantic grid maps are built using traditional sensor fusion methods, then initial map estimation is obtained, but errors from sensor inaccuracies and ambiguities persist

Engineering Contradiction:
Improveease of map buildingVSAvoidmap reliability
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The patent introduces an intermediary neural network component that mediates between raw sensor data and the final map representation. The network learns to interpret ambiguous sensor readings and resolve inconsistencies by comparing multiple observations with the map model, thereby improving map reliability while maintaining the ease of automated map building without requiring manual expert intervention.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4295115B1System and method for training neural network for geographical information
Publication Date: 2025.05.28 CONTINENTAL AUTOMOTIVE TECHNOLOGIES GMBH
  • EP4295115B1 patent drawingFigure 1
  • EP4295115B1 patent drawingFigure 2
  • EP4295115B1 patent drawingFigure 3

AI summary

The present invention relates to a system and a method for training a neural network for geographical information, and particularly relates to a system and a method for updating geographical information using the neural network. In accordance with an aspect of the present invention, there is a system for training a neural network for geographical information comprising: a processor operable to generate an environment from actual geographical information, and simulate a robot moving in the environment to obtain first data from the simulation of the robot, characterised in that: the processor is operable to produce corrupted geographical information from the actual geographical information to obtain second data from the corrupted geographical information, and the neural network is operable to construct predicted geographical information based on the first and second data, and compare the actual geographical information and the predicted geographical information to train the neural network.