Session: 16-01-01: Government Agency Student Poster Competition

Paper Number: 150534

150534 - Tool Grasping Compliance and Stability of Underactuated Hands 

Tendon-driven underactuated hands have become appealing for several reasons, including simplified control schemes, inherent compliance that adapts to unknown shapes during grasping, and low-cost, lightweight designs suitable for scalable use. These hands also benefit from a reduced number of actuators, which leads to lower power consumption and maintenance costs, making them ideal for a wide range of daily and industrial applications. These advantages contribute to the enhanced robustness of underactuated hands compared to fully actuated counterparts. In this study, we aim to investigate the tool grasping stability of underactuated robotic hands and exam how external force disturbances affect tool grasping.

We propose a novel method for estimating tool grasping compliance matrices of underactuated hands by combining feedback from hand joint sensors and wrist-mounted Force/Torque sensors. By leveraging these sensor inputs, we develop a comprehensive model of the underactuated hand as a parallel robot during tool grasping. This model allows us to describe the tool grasping compliance as the relationship between variation in the external wrench applied to the object and the resulting object (moving platform, in the content of parallel robot) deflection in Cartesian space.

Formulating the compliance matrices for an underactuated parallel structure remains challenging due to its nonlinear nature. The nonlinear nature arises from the interaction among the tendons, restoring torsional springs at joints, and external loads. These factors lead to a compliance characteristic that is not only nonlinear but also varies with the configuration and the state of the hand, making traditional stiffness analysis methods insufficient. To address these challenges, we introduce data-driven and machine learning methods to estimate the compliance matrices. These methods involve training models on sensor data to predict the hand's response to external forces and torques.

To effectively generate synthetic data for training our learning models, we have developed a simulation environment that recreates the robotic hand grasping objects of various shapes and sizes. This environment accurately simulates the robotic hand motion as well as provides data-rich interactions at the contact points with objects. Additionally, it allows us to grasp at an object with different grasping configurations and apply disturbance forces at the object. By rapidly running these simulations, we can generate a large volume of data samples. This synthetic data is crucial for training our neural networks, which in turn improves the accuracy of our compliance matrix estimates.

Preliminary results from our data-driven and learning-based methods demonstrate their effectiveness in predicting tool grasping status and stability, highlighting their potential in enabling complex manipulation tasks with the grasped tool.

Presenting Author: Qianwen Zhao Stevens Institute of Technology

Presenting Author Biography: Qianwen Zhao is currently pursuing her PhD in Mechanical Engineering at Stevens Institute of Technology, where she also completed her undergraduate and master's degrees. She earned her Bachelor's degree in Electrical Engineering with a concentration in Robotics and Control in May 2019, followed by her Master's degree in the same field in May 2020.

She is a PhD student in the Advanced Robot Manipulators Lab (ARM lab), under supervision of Dr. Long Wang. Her research primarily explores the dynamics of underactuated robotic hands, specializing in compliance/stiffness analysis, as well as simulations of physical human-robot interactions. Her work seeks to advance the development of safe and effective robotic systems.

Authors:

Qianwen Zhao Stevens Institute of Technology
Long Wang Stevens Institute of Technology