Machine Learning Robot Gripping Control for Soft Objects
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional control methods for gripping objects with small reaction forces, such as soft objects like tofu and cream puffs, fail to accurately detect the necessary reaction force, leading to potential damage due to over-gripping and shape errors.
Innovation Solution
A control device employing machine learning to estimate the gripping width of a robot's hand based on object shape data, using a length measuring sensor instead of force sensors, to securely grip objects with small reaction forces without damaging them.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a force sensor is used to detect reaction force for gripping control, then gripping control can be performed for objects with sufficient reaction force, but gripping control cannot be performed for objects with small reaction force such as soft objects
Solution Approach 1:
The patent replaces the force sensor-based mechanical detection system with a machine learning-based estimation system. Instead of directly measuring reaction force through mechanical contact, the system uses a length measuring sensor to capture object shape data and employs a machine learning model to estimate the appropriate gripping width, thereby eliminating the need for force sensors and enabling control of soft objects with minimal reaction forces.
Solution Approach 2:
The patent introduces a machine learning model as an intermediary between the length measuring sensor and the gripping control system. This intermediary processes shape data and translates it into estimated gripping width parameters, enabling the system to infer appropriate gripping forces without direct force measurement, thus bridging the gap between shape detection and safe gripping control.
2Ease of operation
If gripping control is performed based on motor current and force sensor feedback, then gripping control can be realized for objects with sufficient reaction force, but objects with large shape errors may be too strongly gripped and damaged
Solution Approach 1:
The patent performs preliminary estimation of the appropriate gripping width before actual gripping occurs. The machine learning model predicts the optimal gripping parameters based on pre-captured shape data, allowing the system to prepare safe gripping parameters in advance. This preliminary estimation prevents over-gripping by establishing appropriate force limits before contact is made with the object.
Solution Approach 2:
The patent replaces the traditional feedback-based mechanical control system with a predictive machine learning system. Instead of relying on real-time force feedback during gripping, the system uses shape-based prediction to determine appropriate gripping parameters beforehand, eliminating the need for force sensors and reducing the risk of object damage through more precise, pre-calculated gripping control.
3Reliability
If a force sensor is used for detecting power and moment, then feedback control can be performed, but the system becomes complex and cannot handle objects with small reaction force
Solution Approach 1:
The patent extracts and removes the force sensor component from the sensing system. By eliminating the force sensor, the system simplifies the hardware configuration while maintaining control capability through machine learning-based estimation. The complex force detection and feedback mechanism is replaced with a simpler shape-based prediction approach that achieves comparable or superior performance for soft objects.
Solution Approach 2:
The patent substitutes the mechanical force sensing system with an information-processing-based machine learning system. Instead of using physical force sensors to detect power and moment, the system uses a length measuring sensor combined with machine learning algorithms to estimate gripping parameters, thereby reducing device complexity while maintaining or improving control reliability for objects with small reaction forces.
Data Source
AI summary
A control device and a machine learning device enable control for gripping an object having small reaction force. The machine learning device included in the control device includes a state observation unit that observes gripping object shape data related to a shape of the gripping object as a state variable representing a current state of an environment, a label data acquisition unit that acquires gripping width data, which represents a width of the hand of the robot in gripping the gripping object, as label data, and a learning unit that performs learning by using the state variable and the label data in a manner to associate the gripping object shape data with the gripping width data.