Created: November 16, 2022
Tags: Simulation, Gelsight, Photometric Stereo, PyBullet
Link: https://arxiv.org/pdf/2109.04027.pdf
Status: Reading
What?
A new example-based realistic looking optical and marker field movement simulation for Gelsight-like tactile sensors.
Why?
Previous tactile simulators such as TACTO, Phong’s model lacks the real sensor RGB data and noise. Taxim solves this by including photometric stereo based calibration and shadow modeling
How?
Pipeline:
- Build a polynomial function that simulate the optical response of Gelsight by mapping geometries to pixel intensity in tactile images
- Simulate the shadow effect by dividing objects into small pins and combining individual pin shadows by superposition principle (adding). Shadow simulation is important because real sensor images contain shadows and it will effect optical response.
- Calibration requires less than 100 images of a sphere ball pressed on gel surface.
Limitations
- Optical response errors increase when the indentation become farther from the center and deeper. Authors did not explain why this happens. My personal guess is that 2nd degree polynomial function fails to model the optical response as the gel doesn’t deform after some threshold and the stress-strain relationship enter non-linear region.
- Currently, Taxim only supports quasi-static contact and we need to also simulate dynamic phenomena such as slip
- The simulation pipeline can be improved by GPU acceleration