Tactile-GAT: tactile graph attention networks for robot tactile perception classification

Area of Science:

• Robotics
• Artificial Intelligence
• Sensor Technology

Background:

• Tactile perception is crucial for robots to navigate and interact with complex environments.
• Current deep learning methods often overlook the spatial relationships inherent in multi-channel tactile sensor data.
• Processing simultaneous signals like pressure and bending presents a significant challenge.

Purpose of the Study:

• To develop a novel tactile perception framework for robots.
• To effectively integrate spatial information from tactile sensors into deep learning models.
• To improve the accuracy and robustness of robotic tactile sensing.

Main Methods:

• Proposed a tactile perception framework utilizing graph attention networks (GANs).
• Incorporated both explicit and latent relation graphs to model sensor interdependencies.
• Constructed a tactile glove and collected a dataset of pressure and bending signals during object manipulation.

Main Results:

• Achieved 89.58% accuracy in object tactile signal classification.
• Demonstrated superior performance compared to traditional time-series classification algorithms.
• Showcased the framework's ability to leverage sensor spatial information effectively.

Conclusions:

• The proposed GAN-based framework significantly enhances robotic tactile perception.
• This approach is well-suited for processing complex, multi-channel tactile data.
• The method offers a general strategy for improving robot environmental interaction and decision-making.