A learning-based tip contact force estimation method for tendon-driven continuum manipulator.

Area of Science:

• Robotics
• Machine Learning
• Sensor Technology

Background:

• Tendon-driven continuum manipulators are widely studied but lack efficient tip contact force sensing.
• Existing methods often increase manipulator diameter or rely on complex modeling.
• A general and efficient solution for real-time force sensing is needed.

Purpose of the Study:

• To propose a general tip contact force-sensing method for continuum manipulators.
• To develop a real-time contact force estimator using machine learning.
• To compare the performance of recurrent neural network (RNN) based estimators against a baseline.

Main Methods:

• A recurrent neural network (RNN) model was developed, taking tendon position and tension as input.
• A three-degree-of-freedom (3-DOF) contact force data acquisition platform was designed and built.
• Machine learning was used to train the RNN model for static model fitting and real-time estimation.
• Performance comparison included a multi-layer perceptron (MLP) baseline and three RNN architectures using TensorFlow.

Main Results:

• The proposed RNN-based method demonstrated feasibility for real-time tip contact force estimation.
• Performance of different RNN architectures was evaluated against an MLP baseline.
• A manually decoupled sub-estimators algorithm was proposed and compared with the main method.

Conclusions:

• The RNN-based approach offers a general and efficient solution for tip contact force sensing in continuum manipulators.
• This method avoids increasing manipulator diameter, addressing a key challenge in the field.
• The study validates the use of machine learning for real-time force estimation in robotic manipulators.