Robot Insertion Neural Network Training With Plane-Referenced Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing robot control methods for insertion tasks, such as peg-in-hole, are limited to simple shapes and fixed locations, and visual techniques are slow, typically three times slower than human operators, making efficient training methods for robots in complex assembly processes desirable.
Innovation Solution
A method for training a neural network to derive a movement vector from camera images mounted on a robot to insert an object into an insertion, using a plane as a reference to reduce ambiguity and allowing for the collection of training data offline, which enables the robot to generalize to variations in the control environment, including errors in grasping and different shapes and colors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If existing visual techniques are used for robot insertion tasks, then the robot can perform insertion operations, but the speed is three times slower than human operators
Solution Approach 1:
The patent replaces traditional mechanical vision-based control systems with a neural network-based system. The neural network is trained to directly map camera images to movement vectors, eliminating the need for complex visual processing and geometric calculations that slow down traditional systems. This substitution enables real-time processing at human-like speeds while maintaining adaptability to variations through the neural network's learned representations.
Solution Approach 2:
The patent performs preliminary training of the neural network offline using a dataset of images and corresponding movement vectors. This preliminary action allows the system to learn optimal insertion strategies beforehand, so that during actual operation, the robot can quickly execute insertions by simply querying the trained network with minimal real-time computation, achieving both speed and adaptability.
2Adaptability or versatility
If traditional robot control schemes are used for insertion tasks, then simple shapes in fixed locations can be handled, but the system cannot generalize to variations in shape, color, or location
Solution Approach 1:
The patent replaces complex rule-based control systems with a neural network that learns patterns directly from data. Instead of programming explicit rules for handling various shapes, colors, and locations, the neural network automatically learns invariant features and relationships from training examples, enabling generalization to unseen variations without increasing control system complexity.
Solution Approach 2:
The patent uses a camera to capture visual information and creates a digital representation (image) of the physical insertion target. The neural network processes this copy (image data) to determine movement vectors, allowing the system to generalize across different physical instances without needing to physically interact with each variation during training. This copying approach enables versatile handling of variations while keeping the physical robot system relatively simple.
3Reliability
If training data is collected in real operational environments, then realistic scenarios are captured, but safety hazards may occur during data collection
Solution Approach 1:
The patent collects training data by capturing images of insertion targets in controlled or simulated environments rather than requiring the robot to physically attempt insertions in hazardous real-world scenarios. The neural network learns from these visual copies without the robot needing to physically interact with potentially dangerous objects, eliminating safety hazards while maintaining training data quality through realistic image capture.
Solution Approach 2:
The patent performs data collection and neural network training as a preliminary action before deploying the robot to actual operational environments. By completing the learning phase in advance under controlled conditions, the system avoids safety hazards that would occur during real-time operation, while still capturing realistic insertion scenarios for effective training. The trained model is then deployed for safe operation in production environments.
Data Source
AI summary
A method for training a neural network to derive, from an image of a camera mounted on a robot, a movement vector to insert an object into an insertion. The method includes, for a plurality of positions in which the object held by the robot touches a plane in which the insertion is located controlling the robot to move to the position, taking a camera image by the camera and labelling the camera image with a movement vector between the position and the insertion in the plane and training the neural network using the labelled camera images.