Grasp simulator based on IPC (https://ipc-sim.github.io/), contains modifications from it.
- unzip data.zip, will create
dexgrasp_data/directory populated with mesh + json related data - build using
docker/build.sh - run using
docker/run.sh <obj_name> <grasp_ind>
So, for example, (including setup) you should be able to simulate a cube grasp with:
unzip data.zip
./docker/build.sh
./docker/run.sh cube_3cm 0
This will create grasp state files (.obj files including gripper + object) in output/cube_3cm_0_<time>/.
The grasps can also be visualized at each timestep using any mesh visualizer, but it might be finicky to play through the various timesteps.
If you have blender, then you can
- use https://github.com/neverhood311/Stop-motion-OBJ to load in the object files in order
(the
testprefix can be used to load in the objects). - use the existing
render_grasp.pyfile to create high-quality images of each grasp
python augment_objs.py <output_dir> # render gripper head + etc for improved vis
blender -b -P render_grasp.py <output_dir> # actual render vis, assumes augment_objs already run
In that case, images will be generated in <output_dir>/images.
Can use ffmpeg to stitch them together:
ffmpeg -f image2 -framerate 30 -pattern_type glob -i '<output_dir>/images/*.png' ipc.mp4
ffmpeg -i ipc.mp4 -framerate 5 -vf "split[s0][s1];[s0]palettegen[p];[s1][p]paletteuse" ipc.gif