Robot Controller Generation Using Data-Driven Basis Functions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current robot controller design processes rely on complex and potentially inaccurate manual calculations of kinematic equations, which can lead to inaccuracies in the Jacobian matrix and hinder precise robot control, especially when the physical structure of the robot is hidden or cannot be analyzed.
Innovation Solution
A data-driven approach using a model generator that creates a set of basis functions and coefficients to describe a robot's physical properties, enabling the generation of accurate robot controllers without requiring kinematic equations, and allowing control even when the physical structure is unknown.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual calculation of kinematic equations is used to design robot controllers, then the controller can be generated with knowledge of physical structure, but the modeling assumptions lead to inaccuracies in the Jacobian matrix and control precision
Solution Approach 1:
The patent replaces the traditional mechanical approach of manually deriving kinematic equations and computing Jacobian matrices with a data-driven machine learning system. The system learns the robot's dynamics directly from observed motion data, substituting complex analytical mechanics with empirical pattern recognition that automatically captures the true physical behavior without requiring manual modeling assumptions.
Solution Approach 2:
Instead of creating analytical models that approximate the robot's physics, the system creates a data-driven copy of the robot's actual behavior by learning from observed input-output pairs. This empirical model copies the true dynamics directly from data, avoiding the inaccuracies that arise when analysts must simplify complex physical systems through manual modeling assumptions.
2Ease of operation
If the physical structure of the robot is analyzed to determine kinematic equations, then the controller can be generated, but the physical structure may be hidden and inaccessible to the designer
Solution Approach 1:
The system enables the robot to serve itself in the model generation process by learning its own dynamics from its own operational data. The robot provides its own training data through normal operation, and the system automatically extracts the dynamic model without requiring external analysts to physically inspect or measure the robot's hidden components. The robot essentially teaches itself how to control itself through data-driven learning.
Solution Approach 2:
The patent introduces motion capture data and machine learning algorithms as intermediaries between the robot's physical structure and the controller design process. Instead of directly observing hidden physical components, the system uses observed motion data as an intermediary representation that captures the essence of the physical structure's behavior, allowing controller generation without direct physical inspection.
3Ease of manufacture
If modeling approximations are used to simplify kinematic equations, then the design process becomes more manageable, but inaccuracies in the Jacobian matrix prevent accurate robot control
Solution Approach 1:
The patent fundamentally changes the parameters and methodology of controller design by transitioning from analytical parameters (kinematic equations, Jacobian matrices) to data-driven parameters (learned dynamics models from motion data). This parameter change allows the system to capture complex nonlinear dynamics accurately without requiring simplifying assumptions, as the learning process naturally adapts to the true system behavior from observed data.
Solution Approach 2:
The system replaces the analytical mechanics approach with a data-driven machine learning approach that automatically captures the robot's true dynamics. This substitution eliminates the need for manual modeling approximations entirely, as the learning algorithm directly infers the accurate dynamic model from observed motion data, achieving both ease of design and high control accuracy simultaneously.
Data Source
AI summary
A model generator implements a data-driven approach to generating a robot model that describes one or more physical properties of a robot. The model generator generates a set of basis functions that generically describes a range of physical properties of a wide range of systems. The model generator then generates a set of coefficients corresponding to the set of basis functions based on one or more commands issued to the robot, one or more corresponding end effector positions implemented by the robot, and a sparsity constraint. The model generator generates the robot model by combining the set of basis functions with the set of coefficients. In doing so, the model generator disables specific basis functions that do not describe physical properties associated with the robot. The robot model can subsequently be used within a robot controller to generate commands for controlling the robot.