Robot Depth Image Training for Real-World Sensor Imperfections

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Machine learning models trained on simulated depth data often fail to generalize well to real-world scenarios due to the inability of simulated data to capture imperfections in real-world depth measurements, leading to poor performance in robotics tasks.

Innovation Solution

A method for generating training data that degrades synthetic depth images by replacing depth values at locations with high gradients, curvature, and polygonal patches with invalid values, mimicking real-world data imperfections, allowing the model to better handle real-world sensor data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If simulated depth images are used for training machine learning models, then unlimited training data can be generated, but the models fail to generalize to real-world scenarios due to idealized simulated data not capturing sensor imperfections

Engineering Contradiction:
Improveamount of training dataVSAvoidgeneralization performance
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent applies degradation operations to synthetic depth images to intentionally introduce imperfections such as noise, missing values, and artifacts that mimic real-world sensor limitations. By converting the idealized nature of simulated data into a beneficial training condition, the model learns to handle real sensor imperfections and achieves better generalization performance.

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

Solution Approach 2:

The patent transforms the quality parameters of synthetic depth images by applying various degradation operations including adding noise, creating holes, and introducing artifacts. These parameter changes make the synthetic training data more representative of real-world conditions while maintaining the advantage of unlimited data generation.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If degradation operations are applied to synthetic depth images to match real-world imperfections, then model generalization improves, but the complexity of data generation increases

Engineering Contradiction:
Improvegeneralization performanceVSAvoiddata generation complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent divides the degradation process into multiple independent operations that can be applied separately to synthetic depth images. Each degradation operation (noise addition, hole creation, artifact introduction) is a distinct module that can be individually controlled and combined, making the overall process manageable and systematic rather than monolithic.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies degradation operations during the training data generation phase before the model training begins. By pre-degrading the synthetic depth images, the system prepares training data that already incorporates real-world imperfections, eliminating the need for complex real-world data collection and processing while achieving good generalization.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4238714A1Device and method for controlling a robot
Publication Date: 2023.09.06 ROBERT BOSCH GMBH
  • EP4238714A1 patent drawingFigure 1
  • EP4238714A1 patent drawingFigure 2
  • EP4238714A1 patent drawingFigure 3

AI summary

According to various embodiments, a method for generating training data for a machine-learning model is described comprising obtaining a plurality of depth images, degrading each generated depth image by at least one of replacing depth values at locations with a high depth gradient is above a predetermined gradient threshold by invalid depth values, replacing depth values at locations where an object shown in the depth image has a curvature above a predetermined curvature threshold by invalid depth values and replacing depth values of one or more polygonal patches of the depth image by random depth values or invalid depth values and generating, for each depth image, a training data element comprising the degraded generated depth image as training input data.