Robot Pick-Out Learning Using Workpiece Influence Rate
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing robot systems that learn to pick out workpieces from containers with randomly stacked items often inadvertently damage untargeted workpieces due to the impact force or collapse of the stack during the pick-out action, as they do not effectively account for the influence on untargeted pieces.
Innovation Solution
A machine learning device that observes the behavior patterns of a robot's pick-out action and measures the displaced amount of untargeted workpieces using image data before and after the action, learning an influence rate associated with each behavior pattern to optimize the pick-out action and minimize damage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the robot performs pick-out action to retrieve targeted workpiece, then the workpiece retrieval efficiency is improved, but untargeted workpieces may be damaged due to impact force or stack collapse
Solution Approach 1:
The system performs preliminary observation of the container state and calculates influence rates for different pick-out actions before execution. By evaluating potential impacts on untargeted workpieces in advance and selecting actions with lower influence rates, the system prevents damage before it occurs while maintaining efficient retrieval operations
Solution Approach 2:
The system observes the actual state of workpieces after pick-out actions and uses this feedback to update the influence rate model. By continuously learning from observed displacements and damage patterns, the system refines its prediction accuracy and optimizes future pick-out actions to minimize harm to untargeted workpieces
2Device complexity
If the robot uses conventional supervised data including success/unsuccess results, then the learning process is simplified, but the learned action is far from optimum and may cause damage to untargeted workpieces
Solution Approach 1:
The system introduces a new parameter - influence rate - that quantifies the potential impact of pick-out actions on untargeted workpieces. By incorporating this parameter into the supervised data alongside traditional success/unsuccess outcomes, the system creates a more comprehensive learning objective that guides the robot toward actions that are both successful and minimally harmful
Solution Approach 2:
The system adds a new dimension to the learning problem by considering not just whether a pick-out action succeeds, but also its influence on untargeted workpieces. This transforms the learning objective from a single-dimension success metric to a multi-dimensional optimization that balances retrieval efficiency with protection of remaining workpieces
Data Source
AI summary
An action of a robot included in a robot system is controlled based on a learning result of a machine learning device. The machine learning device includes a behavior observing unit, a displaced amount observing unit, and a learning unit. The behavior observing unit observes a behavior pattern of the robot picking out a workpiece from a container. The displaced amount observing unit observes, based on pieces of image data captured before and after the pick-out action of the robot and output from an image capturing device, a workpiece displaced amount indicating a displaced amount of an untargeted workpiece in the container caused by picking out a targeted workpiece from the container. The learning unit learns an influence rate on the untargeted workpiece with an associated behavior pattern of the robot for picking out the targeted workpiece from the container, the influence rate depending on the workpiece displaced amount.