CHSEL Experiments

This is the official experiments code for the paper CHSEL: Producing Diverse Plausible Pose Estimates from Contact and Free Space Data. If you use it, please cite

@inproceedings{zhong2023chsel,
  title={CHSEL: Producing Diverse Plausible Pose Estimates from Contact and Free Space Data},
  author={Zhong, Sheng and Fazeli, Nima and Berenson, Dmitry},
  booktitle={Robotics science and systems},
  year={2023}
}

Installation (experiments)

1. install pytorch3d (various ways, but probably easiest through conda)
2. install base experiments by following its readme
3. install stucco experiments by following its readme
4. clone repository locally and cd into it
5. pip install -e .

Links

• arxiv
• website

Usage

This is the full experiments to reproduce the results from the paper. See the light-weight library repository for how to use CHSEL in your projects. See the website for videos and a high level introduction.

Registration Experiments

The instructions below are for all methods across all tasks. To specify a set of tasks, use the --task argument, such as --task drill mustard to run only the methods on the drill and mustard pokes. To specify which methods to run, use the --registration argument, such as --registration icp medial-constraint to run only the ICP and medial constraint baselines. By default, 5 random seeds (0,1,2,3,4) are used; to run using other random seeds use the --seed argument, such as --seed 23 42 to run with seeds 23 and 42.

Generate and export data for offline baselines:

python run_many_registration_experiments.py --experiment build --no_gui
python run_many_registration_experiments.py --registration none --no_gui

Generate plausible set for plausible diversity evaluation

python run_many_registration_experiments.py --experiment generate-plausible-set --seed 0 --no_gui

Run poking experiments for all methods (CVO requires preprocessing; see below)

python run_many_registration_experiments.py --experiment poke --no_gui

Evaluate all methods on their plausible diversity

python run_many_registration_experiments.py --experiment evaluate-plausible-diversity --no_gui

Plotting results (images saved under data/img)

python run_many_registration_experiments.py --experiment plot-poke-pd --no_gui

Generate gifs from the logged images after cding into their log directories:

ffmpeg -i %d.png -vf palettegen palette.png
ffmpeg -i %d.png -i palette.png -lavfi paletteuse all.gif

Running Baselines

CVO

1. download docker image https://github.com/UMich-CURLY/docker_images/tree/master/cvo_gpu
2. build docker image and follow instructions
3. first start container with bash run_cuda_docker.bash cvo in the docker/images/cvo_gpu directory (script modified to mount shared data directory)
4. for later uses, restart latest container with "docker start -a -i docker ps -q -l"
5. build CVO
6. run script inside build on a single trajectory bin/cvo_align_manip_freespace ../data/poke/MUSTARD_0.txt ../data/poke/MUSTARD.txt ../cvo_params/cvo_geometric_params_gpu.yaml
7. run script for all trajectories of a task python3 ../scripts/run_many_manip_experiments.py --task mustard mustard_fallen drill