Reproducing Experiments

To reproduce results from our RSS '16 paper, first checkout the RSS '16 version of the codebase:

git checkout tags/v1.0 
cd ~/my_workspace
catkin_make -DCMAKE_BUILD_TYPE=Release

First, download the experiment data:

roscd sbpl_perception 
chmod +x data/scripts/download_experiment_data.sh
./data/scripts/download_experiment_data.sh

This will download the test scenes to sbpl_perception/data/experiment_input and the precomputed RCNN heuristics to sbpl_perception/heuristics. If you haven't already downloaded the object models (for the demo), then run:

chmod +x data/scripts/download_demo_models.sh
./data/scripts/download_demo_models.sh

Finally, run the experiments script:

./experiments/scripts/run_experiments.sh

This will generate results for a run of lazy D2P with 8 processors for parallelization and other default parameters in the run_experiments.sh script. Note that this can take quite some time to run, depending on your machine specifications.

Statistical output will be saved to sbpl_perception/experiments/results_<timestamp>/ and visualization output (output depth images) will be saved to sbpl_perception/visualization/perch_poses_<config_parameters>/. You can modify the settings in run_experiments.sh to sweep across several combinations of parameters including lazy/non-lazy, use RCNN-heuristics/no heuristics, planner suboptimality bound, search resolution, maximum ICP iterations, and number of processors available for parallelization.