Vision-based tactile sensor design using physically based rendering.

Area of Science:

• Robotics
• Sensor Technology
• Computational Optics

Background:

• High-resolution tactile sensors are crucial for robotic fine-grained perception and manipulation.
• Designing these sensors is challenging due to the complex integration of optical elements and difficulty in correlating design with accuracy through trial and error.

Purpose of the Study:

• To introduce a digital design framework for vision-based tactile sensors leveraging physically accurate light simulation.
• To enable systematic evaluation and optimization of tactile sensor designs by quantifying the impact of various design parameters.

Main Methods:

• Development of a modular design framework with parameterized sensor components and a performance evaluation metric.
• Utilizing a physically accurate light simulator to replicate tactile sensor imaging without prior sensor-specific data.
• Quantifying the effects of sensor shape, illumination, and sensing surface material on performance.

Main Results:

• The simulation framework accurately replicates real-world tactile images.
• An improved fingertip GelSight sensor design achieved a 5x performance increase in tactile embossed text detection compared to state-of-the-art human designs.
• The approach is applicable to any vision-based tactile sensor.

Conclusions:

• The digital optical simulation framework facilitates the automatic design of sensorized soft robots.
• This work paves the way for closed-loop co-optimization of controllers and sensors for advanced robotic manipulation.