Calibrating Virtual Physics Engine for Deep Learning Transfer

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep learning-based devices, such as autonomous vehicles, face challenges in learning optimal actions due to a significant gap between virtual and real-world environments, leading to suboptimal performance in real-world scenarios when trained in virtual worlds.

Innovation Solution

A method and device for calibrating the physics engine of a virtual world simulator using a real state network to minimize the gap between virtual and real-world environments by generating predicted next frame information and optimizing physics engine parameters based on the difference between virtual and real-world outcomes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If learning is performed in a virtual world using a physics engine, then training efficiency is improved and time is reduced, but the reliability of learning results deteriorates due to the gap between virtual and real environments

Engineering Contradiction:
Improvetraining efficiencyVSAvoidreliability of learning results
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent adjusts physics engine parameters (friction coefficient, gravity, bounce ratio) to match real-world physical characteristics. By changing these parameters, the virtual environment's dynamics are modified to better align with real-world behavior, thereby improving the reliability of learning results while maintaining the efficiency benefits of virtual training.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces the traditional physics engine simulation with a deep neural network-based dynamic model. This neural network is trained on real-world data to predict next states, substituting the mechanical physics engine with an intelligent system that captures real-world dynamics more accurately, thus improving reliability while maintaining virtual training efficiency.

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

2Reliability

If the physics engine parameters are adjusted to match real-world dynamics, then the alignment between virtual and real environments is improved, but the complexity of the system increases

Engineering Contradiction:
Improvealignment between virtual and real environmentsVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent introduces a deep neural network as an intermediary between the physics engine and the learning agent. This neural network acts as a mediator that translates virtual environment states into accurate predictions of real-world outcomes, improving alignment without requiring complex adjustments to the physics engine itself. The intermediary handles the complexity of matching real-world dynamics.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Measurement precision

If a deep neural network is used to predict next states instead of a physics engine, then the accuracy of virtual-to-real transfer is improved, but the computational resources required increase

Engineering Contradiction:
Improveaccuracy of next state predictionVSAvoidcomputational energy consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent performs preliminary training of the deep neural network offline using real-world data before deployment. By pre-training the network to learn accurate state transitions, the system achieves high prediction accuracy during actual use without requiring intensive real-time computation. The computationally expensive training phase is done in advance, reducing energy consumption during operational phases.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3690750B1Method and device for calibrating physics engine of virtual world simulator to be used for learning of deep learning-based device, and a learning method and learning device for real state network used therefor
Publication Date: 2024.08.14 STRADVISION
  • EP3690750B1 patent drawingFigure 1
  • EP3690750B1 patent drawingFigure 2
  • EP3690750B1 patent drawingFigure 3

AI summary

A method for calibrating a physics engine of a virtual world simulator for learning of a deep learning-based device is provided. The method includes steps of a calibrating device (a) if virtual current frame information corresponding to a virtual current state in virtual environment is acquired, (i) transmitting the virtual current frame information to the deep learning-based device to output virtual action information, (ii) transmitting the virtual current frame information and the virtual action information to the physics engine to output virtual next frame information corresponding to the virtual current frame information and the virtual action information, and (iii) transmitting the virtual current frame information and the virtual action information to a real state network learned to output predicted next frame information in response to action in a real environment to output predicted real next frame information; and (b) optimizing the previous calibrated parameters to generate current calibrated parameters.