Inferring Interaction Force from Visual Information without Using Physical Force Sensors.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Force sensors are traditionally used for interaction force estimation.
• Estimating forces via visual cues is challenging due to subtle shape changes.

Purpose of the Study:

• To develop a novel deep learning method for interaction force estimation using only sequential images.
• To overcome the challenge of detecting subtle shape variations for accurate force prediction.

Main Methods:

• A recurrent neural network (RNN)-based deep model was formulated.
• The model utilizes fully-connected layers to capture complex temporal dynamics from visual data.
• Sequential images of objects undergoing deformation were used as input.

Main Results:

• The proposed method successfully estimated interaction forces using only visual information.
• Accurate force predictions were achieved for diverse objects like sponges, PET bottles, human arms, and tubes.
• Predicted forces closely matched measurements from traditional force sensors.

Conclusions:

• Visual-based interaction force estimation is feasible using deep learning.
• The developed recurrent neural network model effectively models temporal dynamics for force prediction.
• This approach offers a promising sensor-free alternative for force estimation in various applications.