Instructions for Reproducing Experiments from the Paper

Table of Contents:

Instructions for Reproducing Experiments from the Paper

Experiment A: Evaluation of DiffCo Score

2D environment

Use the script scripts/test_correlation.py to generate correlation and FCL comparison figures. If you do not provide a dataset, a default one will be autogenerated for you.

python scripts/test_correlation.py correlate

python scripts/test_correlation.py compare

To create a dataset, use the script 2d_data_generation.py to create data for a handful of predefined environments. Use the --help flag to see all the possible environments.

python scripts/2d_data_generation.py --help`

python scripts/2d_data_generation.py --env 3d_halfnarrow --dof 7 --link-length 1

python scripts/2d_data_generation.py --env 1rect_1circle --dof 3 --link-length 2

Once you have the dataset you want, pass it in to the scripts/test_correlation.py script with the --dataset flag.

3D environment

Documentation coming soon.

Experiment C: Trajectory Optimization that Avoids Important Objects

2D environment

Use the script 2d_trajopt.py to generate trajectory optimization plans and figures. If you do not provide a dataset, a default one will be autogenerated for you.

python scripts/2d_trajopt.py

You can also specify custom start and target configurations.

python scripts/2d_trajopt.py --start-cfg 1 0.5 0 --target-cfg -1 -0.5 0

Using the --safety-margin flag, you can assign a bias to one of the classes.

python scripts/2d_trajopt.py --start-cfg 1 0.5 0 --target-cfg -1 -0.5 0 --safety-margin -12 -1.2

You can pass in a custom dataset with the --dataset flag.

python scripts/2d_data_generation.py --env 2class_2 --dof 7 --label-type class --link-length 1
python scripts/2d_trajopt.py -d data/landscape/2d_7dof_2obs_class_2class_2.pt --start-cfg 3.14 0 0 0 0 0 0 --target-cfg 0 0 0 0 0 0 0

The script also generates a video animation.

3D environment

Documentation coming soon.

Experiment D: Trajectory Optimization with Dynamic Objects

2D environment

Use the script active.py to generate trajectories and figures for snapshots of environments with moving obstacles. If a dataset is not provided, one will be autogenerated. You can generate trajectories using either DiffCo or FCL.