Robot Control System Neural Network Force Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current robot systems require costly torque sensors or force sensors for external force detection, which increase installation and maintenance costs and are mechanically restrictive, and model-based methods struggle with accurate friction modeling and force estimation for medical applications.
Innovation Solution
A control system that estimates external force using state information from a robot equipped with an encoder on the output shaft, a backdrivable speed reducer, and a motor, eliminating the need for torque or force sensors by employing a neural network to learn and estimate disturbances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a torque sensor is mounted on each actuator of a robot joint, then external force detection accuracy is improved, but device complexity and cost increase according to the number of joints
Solution Approach 1:
The patent replaces the mechanical torque sensor with a neural network-based estimation system that uses encoders and backdrivable speed reducers to infer external force from motor current and position data, eliminating the need for physical torque sensing components at each joint
Solution Approach 2:
The patent creates a virtual copy of the torque sensor's function through a neural network model that learns to predict external force from observable system states (encoder positions, motor currents, speed reducer backdrivability characteristics), replacing the need for actual physical sensors
2Measurement precision
If a force sensor is mounted at a distal end of the robot, then external force detection is enabled, but external force cannot be detected at other portions and installation is mechanically restricted
Solution Approach 1:
The patent enables the same encoder-speed reducer-encoder configuration to serve multiple detection purposes across different robot portions by using the neural network to model and predict external forces at various locations based on local state information, making the system universally applicable to all joints without requiring location-specific sensors
3Device complexity
If a model-based method of estimating external force without a force sensor is used, then device complexity is reduced, but measurement precision deteriorates due to difficulty in modeling mechanical friction and friction influence
Solution Approach 1:
The patent replaces traditional analytical friction modeling with a data-driven neural network that learns friction characteristics and other disturbances from operational data, substituting complex mechanical models with a flexible computational model that captures non-linear friction behavior without requiring explicit friction models
Solution Approach 2:
The patent implements a feedback mechanism where the neural network continuously refines its external force estimation by processing real-time encoder position, motor current, and speed reducer state information, allowing the system to adapt to varying friction conditions and improve estimation accuracy dynamically
Data Source
AI summary
A control system of an aspect according to the present disclosure includes a robot including an actuator, and an estimation section that estimates external force received by the robot on the basis of state information of the robot, in which the actuator includes an encoder on an output shaft side, a speed reducer that has backdrivability, that is coupled to the encoder on the output shaft side, and that is a load element, a motor coupled to the speed reducer, and an encoder on an input shaft side which encoder is coupled to the motor.