Model-Based 3D Contact Geometry Perception for Visual Tactile Sensor.

Area of Science:

• Robotics
• Computer Vision
• Sensor Technology

Background:

• Tactile sensing is crucial for robot perception but current technologies have limitations.
• Existing visual perception methods for tactile sensing require extensive training data and complex calibration.
• Visual-tactile sensors (VTS) offer a promising alternative for detailed surface perception.

Purpose of the Study:

• To propose a novel model-based depth perceptual scheme for visual-tactile sensors.
• To develop a robust and efficient method for reconstructing contact geometry and texture.
• To validate the VTS's performance in high-quality geometry reconstruction and robotic pose estimation.

Main Methods:

• A model-based depth perceptual scheme mapping image intensity to contact geometry was formulated.
• A tailored fast solver was developed for efficient computation.
• A stereo model was employed to handle non-uniform illumination conditions.
• Hardware calibration requires only a single image, enhancing algorithmic robustness.

Main Results:

• The VTS prototype achieved high-quality geometry reconstruction of contacting shape and texture.
• Root-mean-square error for shape and texture reconstruction was below the sub-millimeter level.
• Pose estimation experiments demonstrated errors less than 3° in spatial orientation and 1mm in translation.

Conclusions:

• The proposed VTS offers a robust and precise solution for tactile sensing in robotics.
• The method enables accurate reconstruction of surface geometry and texture under varying conditions.
• The VTS is effective for applications like robotic pose estimation, achieving high accuracy.