Joint Angle Estimation of a Tendon-Driven Soft Wearable Robot through a Tension and Stroke Measurement

Byungchul Kim, Jiwon Ryu, Kyu-Jin ChoSensors 2020 · DOI

TL;DR

This paper introduces a data-driven method for accurately estimating the joint angles of a soft wearable robot (Exo-Index) by measuring actuator tension and wire stroke. This approach overcomes the challenges of soft material elongation and human variability, enabling precise control for grasping diverse objects.

What Problem Does This Paper Address?

Developing effective wearable robots, especially those interacting with the human body, presents significant challenges. Key among these are:

  • Human Variability: Differences in human anatomy (bone structure, joint stiffness) make it difficult to design a one-size-fits-all robot.
  • Soft Material Challenges: Soft robots, while offering adaptability and comfort, introduce uncertainties due to material elongation under actuation. This elongation complicates the precise estimation of joint positions and forces.
  • Actuation Complexity: Many soft wearable robots use tendon-driven systems for compactness. However, a single tendon often actuates multiple joints, making it impossible to determine individual joint angles solely from tendon stroke. Instead, joint angles depend on a complex interplay of joint stiffness, wire moment arms, and wire tension, which are hard to model accurately.

Existing methods like wearable sensors (e.g., IMUs, strain sensors) have drawbacks such as error accumulation, discomfort, or susceptibility to environmental factors. Visual methods (cameras) can suffer from occlusions. This paper aims to address these issues by proposing a method that doesn't rely on external sensors on the wearer, instead focusing on internal robot measurements.

Methods

The core of this research lies in a data-driven approach to identify both kinematic and stiffness parameters of the soft wearable robot system.

  • Kinematic Identification:
    • The paper uses a Vicon motion capture system to obtain ground truth joint positions.
  • It employs the Product of Exponential (POE) method to model the relationship between joint angles and the fingertip position. This method is particularly useful for describing the complex transformations in articulated systems like the human finger.
  • A numerical method is used to estimate joint centers of rotation and phalange lengths.
  • Stiffness Identification:
    • This involves establishing the relationship between actuator tension, wire stroke, and joint angle.
  • Gaussian Process Regression (GPR) is used to model this complex, non-linear relationship. GPR is a powerful non-parametric Bayesian approach that can capture complex functions and provide uncertainty estimates.
  • The method simultaneously measures wire tension, wire stroke, and joint angle to train the GPR model.

The proposed Exo-Index robot itself is a novel wearable robot designed to assist the human index finger using three tendon-driven actuators.

Datasets

The paper does not explicitly mention pre-existing datasets for training or evaluation. Instead, it describes experiments conducted on a single participant (pilot study) to collect the necessary data.

  • Experimental Data: Data was collected using a Vicon motion capture system for ground truth joint positions and angles, along with motor encoder and load cell data for tension and stroke measurements. This data was synchronized for analysis.

Key Insights and Ideas

  • Data-Driven Dominance: The research strongly advocates for data-driven methods (like GPR) over traditional model-based approaches for soft wearable robots, acknowledging the inherent difficulties in accurately modeling soft material properties and human biomechanics.
  • POE for Kinematics: The application of the Product of Exponential (POE) formalism provides a robust mathematical framework for describing the complex kinematics of the human finger.
  • GPR for Stiffness: Gaussian Process Regression proves effective in capturing the non-linear relationship between actuator inputs (tension, stroke) and robot outputs (joint angles), even with the complexities introduced by soft materials.
  • Single-Finger Focus: The paper posits that focusing actuation on a single, highly dexterous finger (like the index finger) can be more effective for grasping than distributing limited actuators across multiple fingers with coupled motions. This is based on the idea that controlling a single finger's precise motion is more manageable than coordinating multiple fingers with uncertain dynamics.
  • Grasp Strategy Adaptation: The ability to estimate joint angles and stiffness allows the robot to adapt its grasping strategy (wide pinch, narrow pinch, caging) based on object properties, leading to more effective manipulation.

Key Results

  • Kinematic Identification: Linear regression successfully estimated joint positions and phalange lengths with reasonable accuracy, providing a basis for kinematic modeling.
  • Stiffness Estimation: The data-driven method using GPR achieved low RMSE values for joint angle estimation (around 0.03 rad), comparable to other sensor-based methods. This demonstrates the effectiveness of the approach in estimating the relationship between tension and joint angle.
  • Robot Performance: The Exo-Index robot, controlled using the proposed estimation method, was able to perform three distinct grasp types (wide pinch, narrow pinch, caging) and effectively grasp various objects.
  • Workspace Analysis: The actuated workspace achieved by the robot was found to be 64.08% of the spontaneous workspace of a non-disabled finger, indicating significant assistive capability.

Limitations and Future Work

  • Limitations:
    • No Object Recognition: The current system lacks vision sensors, meaning it cannot automatically identify object properties (shape, size, weight) and thus cannot autonomously decide the optimal grasp posture. Grasp selection is user-dependent.
  • User-Specific Calibration: The kinematic and stiffness parameters are user-specific. If human properties change, re-estimation (requiring a Vicon system) is necessary.
  • Practical Issues: The paper notes that the proposed robot system cannot configure object information on its own.
  • Future Work:
    • Integration of RGBD Camera: The authors plan to incorporate an RGBD camera to enable the system to identify object and hand status, and to estimate object properties (size, shape) and hand status (joint angle, position).
  • Learning Algorithm Development: The goal is to develop a learning algorithm that can leverage the camera input for better object and hand state estimation.

Summary updated

BibTeX
@article{kim2020joint,
  author = {Byungchul Kim and Jiwon Ryu and Kyu-Jin Cho},
  title = {{Joint Angle Estimation of a Tendon-Driven Soft Wearable Robot through a Tension and Stroke Measurement}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20102852},
}
Figure