Robot Grasp Detection Learning With Continuous Angle Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for training models to detect gripping positions for picking robots face instability due to varying object orientations and increased calculation complexity when incorporating rotation angles, leading to suboptimal performance.
Innovation Solution
A learning method that generates a first inclined rectangle representing the gripping position, updates the model to minimize errors in rectangle and inclination outputs, allowing for stable training and continuous representation of rotation angles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If bounding boxes with different rotation angles are prepared in the SSD, then the model can detect gripping positions for objects in various orientations, but the calculation amount enormously increases
Solution Approach 1:
The patent changes the parameter representation from discrete rotation angles to continuous sine and cosine values. Instead of preparing multiple bounding boxes with different rotation angles, the model learns to output (sin θ, cos θ) pairs that continuously represent any rotation angle, thereby maintaining adaptability for various object orientations while significantly reducing calculation amount.
2Measurement precision
If the rotation angle is included as one of items to be learned in the SSD, then the model can represent continuous rotation angles, but the rotation angle becomes discrete at the upper limit value and the lower limit value leading to unstable training
Solution Approach 1:
The patent inverts the approach by not directly learning the rotation angle θ, but instead learning (sin θ, cos θ). This inversion avoids the discrete discontinuity problem at angle boundaries because sine and cosine functions are continuous and periodic, naturally handling the wrap-around at 0°/360° without creating artificial discontinuities that disrupt training stability.
3Device complexity
If traditional SSD bounding boxes are used, then the model structure is simple, but the gripping positions cannot be uniquely identified including rotation angles for objects in various orientations
Solution Approach 1:
The patent merges the rotation angle information with the bounding box representation by combining (sin θ, cos θ) outputs with the standard bounding box coordinates. This integration allows the model to simultaneously output both position and orientation information in a unified framework, maintaining relatively simple model structure while achieving accurate gripping position identification for objects in various orientations.
Data Source
AI summary
A learning method is performed by a computers The method includes: receiving a first image that includes an object; generating a first rectangle in the first image, the first rectangle including therein a figure, that is set in advance to have a first inclination and that represents a gripping position of an object, and having a side parallel to a first direction; inputting the first image to a model, which outputs, from the input image, a rectangle parallel to the first direction and an inclination, to cause the model to output a second rectangle and a second inclination; and updating the model such that errors of the second rectangle and the second inclination with respect to the first rectangle and the first inclination respectively decrease.