Robotic Grasping Model Adaptation Across Simulated and Real Domains

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine learning-based approaches for robotic grasping require extensive real-world testing, which is time-consuming, resource-intensive, and wears out physical robots, necessitating a more efficient training method.

Innovation Solution

A machine learning model is trained using simulated data and then adapted with domain-adversarial similarity losses based on both simulated and real-world data, reducing the need for extensive real-world training examples and minimizing robot wear.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If training is performed using only real-world physical robot data, then the model achieves accurate robotic grasping performance, but the training process consumes excessive time, resources, and causes robot wear and tear

Engineering Contradiction:
Improverobotic grasping performanceVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent creates a simulated environment that copies real-world physics and robot mechanics to generate synthetic training data. This virtual copy allows extensive training without physical robot wear, while domain adaptation techniques ensure the simulated data maintains fidelity to real-world scenarios, resolving the contradiction between training efficiency and performance accuracy

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent performs preliminary training in simulation before deploying to real robots. By pre-training the model extensively in the virtual environment and then fine-tuning with limited real data, the approach achieves high performance while minimizing the time and wear associated with real-world training iterations

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If extensive real-world training data is collected, then the model achieves better grasping accuracy, but the process requires heavy human intervention and resource consumption

Engineering Contradiction:
Improvegrasping accuracyVSAvoidhuman intervention requirement
Core Design Contradiction:
Measurement precisionVSEase of operation

Solution Approach 1:

The simulated environment automatically generates training data without human intervention. The system self-creates diverse training scenarios, executes robot actions, and collects results autonomously, eliminating the need for human operators to manually position objects and monitor each training trial while maintaining high data quality through realistic physics simulation

Inventive Principle:
Principle #25Self-service

3Productivity

If simulated training data is used exclusively, then training efficiency improves and robot wear is minimized, but the model performance on real robots deteriorates due to simulation-reality gap

Engineering Contradiction:
Improvetraining efficiencyVSAvoidmodel performance on real robot
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent employs domain adaptation techniques that dynamically adjust model parameters and features to bridge the simulation-reality gap. By learning domain-invariant representations and adapting to real-world distributions, the model maintains high performance on physical robots while being trained primarily on simulated data, resolving the contradiction between training efficiency and transferability

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3624999B1Machine learning methods and apparatus for robotic manipulation and that utilize multi-task domain adaptation
Publication Date: 2024.12.18 GOOGLE LLC
  • EP3624999B1 patent drawingFigure 1A
  • EP3624999B1 patent drawingFigure 1B
  • EP3624999B1 patent drawingFigure 2A

AI summary

Training a machine learning model that, once trained, is used in performance of robotic grasping and/or other manipulation task(s) by a robot. The model can be trained using simulated training examples that are based on simulated data that is based on simulated robot(s) attempting simulated manipulations of various simulated objects. At least portions of the model can also be trained based on real training examples that are based on data from real-world physical robots attempting manipulations of various objects. The simulated training examples can be utilized to train the model to predict an output that can be utilized in a particular task – and the real training examples used to adapt at least a portion of the model to the real-world domain can be tailored to a distinct task. In some implementations, domain-adversarial similarity losses are determined during training, and utilized to regularize at least portion(s) of the model.