Robotic System Simulation With Neural Network Error Correction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing robotic system models are inadequate for simulating complex robotic systems like robotic vehicles, as they fail to provide accurate, stable, and reliable behavior simulations, especially in safety-critical applications where small errors can lead to collisions.

Innovation Solution

A simulator system that combines an approximative model of the robotic system with a trained neural network to correct and enhance the simulation accuracy, using a correction module that includes normalization and PCA transformations to improve the realism of the simulation by predicting the difference between the real-world behavior and the model's output.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If an approximative model is used for simulation, then computational efficiency is improved, but simulation accuracy deteriorates

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidsimulation accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

A correction module acts as an intermediary between the approximative model and the ground truth. This correction module, implemented as a neural network, learns the difference between the approximative model's predictions and actual system behavior, then applies this correction to improve simulation accuracy while maintaining computational efficiency of the underlying approximative model

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system transforms the state space by applying normalization and PCA (Principal Component Analysis) to the state vector. This parameter transformation improves the neural network's ability to learn corrections by presenting normalized, decorrelated inputs, thereby enhancing simulation accuracy without increasing the complexity of the base approximative model

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If a neural network correction module is added, then simulation accuracy is improved, but device complexity increases

Engineering Contradiction:
Improvesimulation accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

Rather than building a complete complex model from scratch, the system uses a lightweight neural network that only learns the correction term (the difference) between the simple approximative model and ground truth. This partial action approach achieves high accuracy with minimal added complexity

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The simulation system is segmented into distinct functional modules: the approximative model for base predictions, the correction module for accuracy improvement, and preprocessing modules (normalization, PCA) for state preparation. This segmentation allows each component to be optimized independently and facilitates training and deployment

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4068033B1Method and device for simulating a robotic system
Publication Date: 2024.07.10 COMMISSARIAT A LENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES
  • EP4068033B1 patent drawingFigure 1~2
  • EP4068033B1 patent drawingFigure 3~4
  • EP4068033B1 patent drawingFigure 5~6

AI summary

The present disclosure relates to a method of simulating the behavior of a robotic system using a computing system comprising a processing device, the method comprising: providing, to an approximative model (304) of the robotic system and to a correction module (306), a state vector (Xk) indicating a state of the robotic system, and a command vector (Uk), wherein the correction module comprises a trained neural network (308); generating, a rate of change vector (ẊMODEL); generating, using the trained neural network and based on the state vector and on the command vector, a correction vector (δẊ) indicating a correction to be applied to the rate of change vector (ẊMODEL); and estimating a next state vector (Xk+1) of the robotic system based on the rate of change vector (ẊMODEL) corrected by the correction vector (δẊ).