Robot Manipulation Training Data With Force, Vision, and Stiffness

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for programming robots to perform dexterous contact-rich manipulations face challenges such as sample efficiency, selecting optimal stiffness parameters, and teaching interfaces, particularly in controlling rigid robots like collaborative or industrial robots.

Innovation Solution

A method and entity for obtaining a training data set and controlling rigid robots using a remote controller to provide position, strength, and image information, combined with compliance control, to train a neural network for precise manipulation tasks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If Learning from Demonstrations (LfD) is used to teach robots dexterous manipulations, then the intuitiveness of programming is improved, but the sample efficiency deteriorates (requiring many demonstrations for each task)

Engineering Contradiction:
Improveintuitiveness of programmingVSAvoidnumber of demonstrations required
Core Design Contradiction:
Ease of operationVSLoss of time

Solution Approach 1:

The system performs preliminary action by pre-processing the demonstration data to extract key features and patterns. The teleoperation demonstrations are captured and processed offline to create a compact representation that can be quickly applied during actual robot execution, reducing the need for repeated demonstrations.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system uses copying by creating a virtual model or representation of the demonstrated task from the teleoperation data. Instead of requiring the robot to physically repeat each demonstration, the system captures the essential patterns and copies them into a reusable model that guides the robot's actions, significantly reducing the number of demonstrations needed.

Inventive Principle:
Principle #26Copying

2Adaptability or versatility

If compliance control schemes are applied to allow rigid robots to handle contact tasks, then the ability to perform contact-rich manipulations is improved, but the complexity of selecting optimal stiffness parameters increases

Engineering Contradiction:
Improveability to perform contact-rich manipulationsVSAvoidcomplexity of selecting stiffness parameters
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system implements feedback by using the captured teleoperation data to inform and adjust the compliance control parameters. The demonstrations provide implicit feedback about the appropriate stiffness and compliance settings for different task phases, allowing the system to automatically select optimal parameters without manual tuning complexity.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system applies parameter changes by dynamically adjusting the stiffness and compliance parameters based on the task phase and contact conditions learned from demonstrations. Instead of requiring manual selection of fixed parameters, the system automatically modifies these parameters in response to task requirements, simplifying the overall control complexity while maintaining adaptability.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If a trained control model is used to determine robot positions based on position, strength, and image information, then the control accuracy is improved, but the amount of training data required increases

Engineering Contradiction:
Improvecontrol accuracyVSAvoidamount of training data
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The system extracts only the essential and discriminative features from the teleoperation demonstrations, such as key position landmarks, critical contact force patterns, and important visual cues. By taking out only the most relevant information rather than using complete raw data, the system achieves high control accuracy with a reduced training dataset.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system performs preliminary processing and feature extraction on the demonstration data before training the control model. This preliminary action consolidates the essential information into a compact form, allowing the model to learn effective control policies from fewer demonstrations while maintaining high accuracy.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4667168A1Obtaining training data sets for controlling a robot and controlling a robot based on a trained model
Publication Date: 2025.12.24 OMRON CORP
  • EP4667168A1 patent drawingFigure 1~2
  • EP4667168A1 patent drawingFigure 3
  • EP4667168A1 patent drawingFigure 4

AI summary

A method is provided for obtaining a training data set for controlling a rigid robot to perform manipulation by means of an end effector, EE. The method includes controlling (S10), based on control information provided by a remote controller operated by a user, the robot to perform a manipulation task, the control information indicating at least a remotely controlled position that the robot is to take during the manipulation task according to an operation of the remote controller by the user. Further, the method includes obtaining (S20) a training data set including position information and strength information in association with each other, wherein the position information indicates one or more positions of the EE during the manipulation task in response to the control information and the strength information indicates one or more strength values detected at the EE in correspondence of a respective one of the one or more positions of the EE. The training data further includes, in association with the position information and the strength information, stiffness information and image information, wherein the stiffness information indicates a stiffness applied by the rigid robot in correspondence of the position information, and wherein the image information represents one or more images in correspondence of a respective one of the one or more positions of the EE, wherein preferably each of the one or more images shows at least a portion of the EE.