Awesome

MIT RLG GelSight Driver

This repository includes both calibration and driver code for using a GelSight touch sensor.

Installation and Dependencies

This repository has been tested on Ubuntu 14.04. It probably works on other versions of Linux, at least, but they are not actively supported yet. (One thing at a time.)

For quick setup on Ubuntu, use:

sudo apt-get install git cmake pkg-config libopencv-dev python-opencv xawtv libyaml-cpp-dev libboost-all-dev

pkg-config is necessary for the build, opencv is used for grabbing images from webcams and some image processing, xawtv is used to adjust exposure settings on the GelSight webcam, 'yaml' is used for configuration and calibration files.

The python files also need dependencies:

pip install numpy scipy

and maybe some more that I'm forgetting (please update this README if you find them!).

To build, follow the standard CMake workflow:

mkdir -p build
cd build
cmake ..
make
make install

Installed binaries are placed in build/install/bin.

This repository includes two versions of (WIP!) drivers for the Gelsight sensor. A C++ version capable of reading the webcam (but not doing reconstruction) is in src/gelsight_depth_driver.*. It will eventually be fleshed out, but a Python prototype of the calibration and driver functionality is instead available in files src/do_*.py.

Running a GelSight Sensor

The data flow with a GelSight sensor is:

The webcam inside of the camera takes an RGB image.
The driver maps from RGB to normal vector at each pixel using a calibration profile.
The driver outputs a depth map for external consumption (which, for now, can be done via a render window to observe the pretty pictures, or over LCM).

Running the Depth Driver

Assuming you have a calibration file available (one is provided in src/filtered_2017-08-22T19:03:38.648270.calib, performed on the blue-and-electrical-tape Gelsight unit in 32-380), you can run the Python depth driver with:

python do_normal_and_depth_reconstruction.py <path_to_calibration_file>

This application will show three images side-by-side:

Example Python Driver Image

You can try running build/install/bin/run_gelsight_depth_driver --help to see its command-line options, and use it to get raw webcam images to sanity-check that your Gelsight images look OK.

Generating Calibration Data

To generate your own calibration via the Python scripts, you'll need two things:

A rod with a square cross section of known side-length, in order to calibration the lateral scaling of the sensor (pixels-to-lateral-mms scaling). I use 1/4" square stock, and the parameters in the do_rod_calibration.py script assume this size.
A bearing ball of known size. do_ball_calibration.py assumes a 1/4" ball. This is for calibrating the RGB-to-normal-map step.

The procedure is (and this is still WIP/hacky!):

Use build/install/bin/run_gelsight_depth_driver -s rod to collect images of you touching the 1/4" rod to the surface of the Gelsight.
Run python do_rod_calibration <path to clear rod image> <path to background image> and then click on the four corners of the rod. The script should spit out a `rod_calib_*.calib" file and report a reasonable pixel-to-mm scaling (something like 30mm per pixel).
Use build/install/bin/run_gelsight_depth_driver -s rod to collect images of you rolling the ball bearing on the Gelsight. Collect images of the ball in a few different locations, and a few different pressures (sizes).
Use python do_ball_calibration <path to a clear ball image> <path to a ball bg image> <path to rod calib file> on each clear ball image that you like (I recommend at least 5 or 6). Click on a few points around the edge of the circle until a circle fit appears, then hit any key to exit. It should spit out a ball_*.calib file.
Run python do_process_calibrations.py <first ball calib> <second ball calib> ... <last ball calib>, which should spit out a filtered_*.calib file. This is your calibration for the Gelsight! (It's a YAML file containing the mm_to_pixel scaling, as well as a grid of RGB-to-normal calibration points that the driver uses to perform the reconstruction.