Robot Control Using Relative Object Relations for Versatile Tasks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional robot control methods lack versatility, requiring new learning when environmental or object changes occur, leading to high teaching costs due to direct association of time-series control commands with specific operations.
Innovation Solution
A control device that generates control commands based on relative relationship amounts between objects, allowing for operation execution regardless of detailed changes by associating commands with relative position, posture, and state changes, using sensors like cameras to acquire data and determine control commands through learned models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If time-series control commands are directly associated with specific operations through manual teaching or conventional machine learning, then the robot can execute the taught operation accurately, but the teaching cost increases significantly when components or operations change
Solution Approach 1:
The patent introduces an intermediary representation layer between the robot's control commands and the actual operations. Instead of directly teaching time-series control commands for each specific operation, the system learns a mapping from operation descriptions to control commands through a learned model. This intermediary learning process allows the robot to generalize across different operations and components, reducing teaching costs while maintaining execution accuracy.
2Reliability
If the robot is manually moved to directly teach operations for each component variation, then the operation can be executed correctly for that specific component, but the teaching cost becomes too high to be practical
Solution Approach 1:
The patent uses copying by learning a general mapping relationship from operation descriptions to control commands. Instead of manually teaching each specific operation for every component variation, the system creates a learned model that copies the essential control patterns across different operations. This learned model can generate appropriate control commands for new operations based on their descriptions, significantly reducing the need for repetitive manual teaching while maintaining operation correctness.
3Adaptability or versatility
If conventional machine learning methods are used to automate part of the teaching process, then some teaching costs are reduced, but the robot still requires new learning when environmental or object changes occur
Solution Approach 1:
The patent implements universality by creating a learned model that handles multiple operations and component variations through a single general mapping relationship. The model learns from diverse operations and can generalize to new operations and components without requiring specific retraining for each case. This universal approach allows the robot to adapt to environmental and object changes more efficiently, reducing the time needed for relearning while maintaining teaching process automation.
Data Source
Figure 1
Figure 2A
Figure 2B~2C
AI summary
Provided is a technical feature for enhancing versatility of an ability to execute a task to be learned. A control device according to one aspect of the present invention sets a relative relation amount between a plurality of target objects that are to be final objectives. In addition, the control device repeatedly acquires observation data from a sensor and calculates the relative relation amount between the plurality of target objects existing in the environment from the acquired observation data. Further, the control device determines a series of the relative relation amounts in a state that is to be an objective from the relative relation amount at a time point at which control of the behavior starts until the relative relation amount as the final objectives is realized, and repeatedly determines control instructions so as to change the relative relation amount in a present state calculated from the latest observation data into the relative relation amount in a state of an objective to be transitioned to next until the relative relation amount as the final objective is realized. Then, the control device outputs the determined control instruction to a robot device.