Mobile Manipulator Object Selection Using Relative Position Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for controlling mobile manipulators, which are movable, cannot determine the optimum movement-target object to be moved to a predetermined position effectively due to their mobility, unlike fixed robots, where the robot is positioned relative to target objects and obstacles.
Innovation Solution
A computing device equipped with a trained machine learning model that uses state variables including relative positions of target objects and obstacles to the mobile manipulator, route lengths, and holdable areas to determine the optimal movement-target object, allowing for appropriate decision-making even when the manipulator's absolute position changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a machine learning model uses absolute positions of target objects, then the model can determine movement targets for a fixed robot, but the model cannot appropriately determine movement targets when the robot moves to different positions
Solution Approach 1:
The patent transforms the coordinate system from absolute position to relative position. The state variable contains relative positions of target objects to a specific portion of the mobile manipulator (e.g., manipulator base or end effector) instead of absolute positions. This parameter transformation allows the machine learning model to generalize across different manipulator positions, solving the contradiction between reliability for fixed robots and adaptability for mobile manipulators.
2Measurement precision
If the state variable includes detailed information about all target objects, then the model can make accurate decisions, but the computation amount and training time increase
Solution Approach 1:
The patent extracts only the essential features needed for movement target determination into the state variable. Specifically, it uses relative positions of target objects and obstacles, and represents the environment as a simplified map or bitmap image. This extraction approach maintains sufficient accuracy for decision-making while significantly reducing the computational burden and training time compared to using complete detailed information about all objects.
Solution Approach 2:
The patent creates a simplified representation (copy) of the environment using map data or bitmap images that capture the essential spatial relationships without containing all detailed information about target objects. This copied representation is sufficient for the machine learning model to determine movement targets accurately while reducing computation time and training requirements.
Data Source
AI summary
A computing device performs computation for controlling operations of a mobile manipulator configured to hold a plurality of target objects with a manipulator and move the target objects to predetermined positions. The computing device includes a storage and a calculator. The storage stores a trained machine learning model trained by inputting a plurality of training data sets, which are combinations of state variables and pieces of determination data associated with the state variables. The training data sets are acquired in advance. The calculator outputs a movement-target object to be moved to a predetermined position at current time by inputting the state variable to the trained machine learning model read from the storage. The state variable contains relative positions of the target objects to a specific portion of the mobile manipulator. The determination data associated with the state variable represents the movement-target object.