Dynamic-Vision-Based Force Measurements Using Convolutional Recurrent Neural Networks.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Accurate contact force measurement is crucial for robotic manipulation and safety.
• Existing methods often struggle with varying object sizes and require direct physical contact.
• Dynamic Vision Sensors (DVS) offer high temporal resolution for capturing rapid changes.

Purpose of the Study:

• To propose a novel dynamic vision-based measurement method for contact force estimation.
• To develop a system capable of measuring contact force independent of object sizes.
• To validate the robustness and accuracy of the proposed method across different object sizes.

Main Methods:

• Utilized a neuromorphic camera (Dynamic Vision Sensor) to capture intensity changes in a silicone membrane during contact.
• Developed and implemented three deep Long Short-Term Memory (LSTM) neural networks combined with convolutional layers.
• Trained networks to estimate contact force from spatio-temporal intensity changes over time.

Main Results:

• Demonstrated robustness against variable contact sizes, with networks learning object size early in grasp.
• Achieved accurate contact force estimation every 10 ms using spatial and temporal features.
• Obtained a Mean Squared Error (MSE) of less than 0.1 N for grasping and holding forces via leave-one-out cross-validation.

Conclusions:

• The proposed dynamic vision-based method accurately measures contact force independent of object size.
• Deep learning models, particularly LSTMs with memory gates, are effective for this task.
• The approach shows significant promise for real-time robotic applications requiring precise force sensing.