Robot Camera Skill Transfer Through Manual Motion Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies require extensive learning data and time to maintain imaging quality with robotic systems, and there is a pressing need to transfer skilled worker skills due to aging populations and decreased workforce, making automation challenging.
Innovation Solution
A skill transfer mechanical apparatus that allows manual motion correction by operators, generating and learning from manual corrections to automate tasks efficiently, reducing the need for extensive learning data and time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a robot camera is controlled only by a neural network during automatic control, then the imaging quality can be maintained at the same level as a skilled cameraman, but an enormous amount of learning data and a long start period are required
Solution Approach 1:
The system performs preliminary action by having the operator manually correct the robot camera's operation before automatic control begins. These manual corrections are stored as correction data that the neural network learns from, allowing the system to start with pre-processed knowledge rather than requiring extensive learning from scratch.
Solution Approach 2:
The system creates a copy of the skilled operator's corrections by having the neural network learn from the correction data generated during manual operation. Instead of learning all imaging decisions from scratch, the network copies the correction patterns established by the skilled operator, significantly reducing the learning period while maintaining imaging quality.
2Extent of automation
If a robot camera is controlled only by a neural network, then automation is achieved, but the number of positions of the photographic object is infinite requiring enormous learning data
Solution Approach 1:
The system introduces an intermediary element - the correction data generated during manual operation - that bridges the gap between limited manual corrections and the infinite variety of object positions. The neural network learns from these correction patterns rather than requiring direct examples of all possible positions, enabling automation with reduced learning data.
Solution Approach 2:
The system changes the parameter of learning data representation from raw imaging data to correction data. By transforming the learning task from learning all possible imaging scenarios to learning correction patterns, the volume of required learning data is dramatically reduced while maintaining the ability to handle infinite object positions through generalization.
3Reliability
If extensive learning data is collected to maintain imaging quality, then the neural network can be trained effectively, but the time required for automation increases significantly
Solution Approach 1:
The system extracts only the essential correction patterns from manual operation rather than requiring extensive learning data. By taking out and storing only the correction actions needed to maintain imaging quality, the system achieves reliable automation with minimal data collection time, separating the essential learning content from redundant information.
4Loss of time
If the neural network learns from manual motion corrections, then automation is achieved in a short period, but the operator must perform manual corrections during the learning phase
Solution Approach 1:
The system implements self-service by having the operator's manual corrections automatically captured and used to train the neural network. The corrections made during normal operation are automatically stored and processed, allowing the system to learn from the operator's expertise without requiring separate training sessions or additional operator effort beyond normal operation.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A skill transfer mechanical apparatus (500) is capable of correcting a motion of a working part (217) by an operator physically applying a force to the working part (217) directly or indirectly, and includes an operating part (201) configured to move the working part (217) so as to perform a work, a controller (203) configured to control a motion of the operating part (201), a motion information detector (254) configured to detect motion information of the operating part (201) corresponding to the motion of the working part (21), and a manual motion correcting data generator (282) configured to subtract an automatic motion instruction from motion data included in the motion information detected by the motion information detector (254) to generate manual motion correcting data. The controller (203) includes a basic motion instructing module (250) configured to output a basic motion instruction, a learning module (252) configured to output an automatic motion correcting instruction, a motion correcting data generator (272) configured to add the manual motion correcting data to the automatic motion correcting instruction to generate motion correcting data, a motion correcting data storing module (283) configured to store the motion correcting data, and a motion information storing module (254) configured to store the motion information. The learning module (252) carries out machine learning of the motion correcting data stored in the motion correcting data storing module (283) by using the motion information stored in the motion information storing module (256), and after the machine learning is finished, accepts an input of the motion information during the operation of the operating part (201), and outputs the automatic motion correcting instruction, and the operating part (201) moves the working part (217) according to the automatic motion instruction based on the basic motion instruction and the automatic motion correcting instruction, and the manual motion correction.