Robot Action Transfer Using Few-Shot SOM Adaptation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current action transfer techniques require a large number of learning samples from both the transfer source and destination robots, making the process time-consuming and costly, especially when the transfer destination robot is a real robot in an actual operation setting, where acquiring matching samples is difficult.

Innovation Solution

An action transfer method using the Self Organizing Map (SOM) algorithm, which acquires and corrects action information from a transfer source robot using a smaller number of samples from a transfer destination robot, adapting joint values and coordinate values through an update formula to generate transferable action data, even when physical properties are unknown.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a large number of learning samples are acquired from both transfer source and destination robots, then action transfer accuracy is improved, but the time and cost required for action transfer increases significantly

Engineering Contradiction:
Improveaction transfer accuracyVSAvoidaction transfer time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary action by acquiring a large number of learning samples from the transfer source robot in advance, before the actual action transfer to the destination robot. This pre-acquired sample set is then used to generate transfer data, eliminating the need to acquire equivalent samples from the destination robot and significantly reducing the time required for action transfer while maintaining accuracy.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If physical properties of the transfer destination robot are measured and used for action information processing, then adaptation accuracy is improved, but the complexity of advance preparation increases

Engineering Contradiction:
Improveadaptation accuracyVSAvoidadvance preparation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent creates a virtual model (copy) of the transfer destination robot's physical properties, including link lengths, joint ranges, and inertia parameters. This virtual model is then used to process and adapt action information from the transfer source robot, eliminating the need for complex physical measurement and calibration procedures while achieving accurate adaptation through simulation-based processing.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3536467B1Action transfer device, action transfer method, and non-temporary computer readable medium having action transfer program stored thereon
Publication Date: 2022.07.20 SOINN INC
  • EP3536467B1 patent drawingFigure 1
  • EP3536467B1 patent drawingFigure 2
  • EP3536467B1 patent drawingFigure 3

AI summary

An action transfer from a transfer source robot is implemented by using a small number of learning samples acquired by a transfer destination robot. A transfer source action information acquisition unit (101) acquires first action information of the transfer source robot. A transfer destination action information acquisition unit (102) acquires second action information of the transfer destination robot. A correction unit (103) corrects the action information of the transfer source robot by using the second action information and in accordance with a predetermined update formula and thereby generates third action information of the transfer destination robot. The pieces of the second action information of the transfer destination robot are less than the pieces of the first action information of the transfer source robot. The first to third action information includes a set of data indicative of one or more robot joint values and a set of data indicative of a coordinate value of a robot specific part.