Neural Network Robot Control Without Inverse Kinematics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing robot control technologies are not suitable for diverse environments and require specialized robotics engineers, as they struggle with accurately controlling robots for tasks like picking up small or undefined objects, such as cloth and liquid, due to limitations in calculating the position and posture of these items and errors in inverse kinematics.
Innovation Solution
A robot controller that uses a neural network updated by virtual images and environmental conditions, allowing it to drive the robot based on a policy learned in a virtual environment and applied in a real environment, without relying on inverse kinematics or operation plans, enabling precise control of a wide range of robots.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If inverse kinematics and operation plan are used to control the robot arm, then the control method is simple and based on geometric calculations, but the positioning accuracy deteriorates due to errors in arm dimensions, rigidity, and measurement
Solution Approach 1:
The patent replaces the mechanical control system based on inverse kinematics with a neural network-based control system. The neural network learns the mapping from image inputs to arm control outputs through training, substituting the geometric calculation approach with a data-driven approach that adapts to actual physical variations in the robot arm, thereby improving positioning accuracy without increasing control complexity
Solution Approach 2:
The patent creates a virtual copy of the robot arm and environment in a virtual space, where the neural network is trained using simulated images and corresponding control outputs. This virtual training environment allows the system to learn accurate control policies without requiring physical trial-and-error, and the learned policy is then transferred to control the real robot arm
2Measurement precision
If a neural network with three layers or more is used to calculate work position and posture from images, then the technique can handle complex visual recognition, but it becomes unsuitable for controlling robots that pick up objects with undefined position and posture such as cloth and liquid
Solution Approach 1:
Instead of trying to calculate the position and posture of objects with undefined characteristics (like cloth and liquid) from images, the patent inverts the approach: the neural network directly outputs control commands for the robot arm based on image inputs, without explicitly calculating object position and posture. This bypasses the limitation of defining object states while still achieving effective manipulation
Solution Approach 2:
The patent extracts only the necessary control information from images without requiring full interpretation of object position and posture. The neural network learns to map image features directly to control outputs, extracting only the essential information needed for successful manipulation while ignoring unnecessary details about object state
3Adaptability or versatility
If deep learning is performed in a virtual environment to generate a convertor for calculating work position and posture, then the technique can be applied without specialized robotics engineers, but it requires complex virtual environment setup and large amounts of training data
Solution Approach 1:
The patent creates a universal neural network controller that can be applied to various robot arms and tasks through a single training process in the virtual environment. The learned policy is general enough to handle different scenarios without requiring specialized engineering knowledge for each application, making the system universally applicable while the complex virtual setup is performed only once during training
Data Source
AI summary
A robot controller that controls a robot by automatically obtaining a controller capable of suitably controlling a wide range of robots. An image is acquired from an image capturing apparatus that photographs an environment including the robot. The robot is driven based on an output result obtained by inputting the image to a neural network. The neural network is updated according to a reward generated in a case where a plurality of virtual images photographed while changing an environmental condition of a virtual environment generated by virtualizing the environment and a state of a virtual robot are input to the neural network, and a policy of the virtual robot, which is output from the neural network, satisfies a predetermined condition.