Repository associated with our IROS 2023 (accepted) paper https://arxiv.org/abs/2310.08270. The repo contains trajectory optimizers for planning under multi-modal uncertainty.
IROS23_2206_VI_i.mp4
- Clone this repository:
git clone https://github.com/arunkumar-singh/RKHS_Stochastic_Traj_Opt.git
cd RKHS_Stochastic_Traj_Opt
- Create a conda environment and install the dependencies:
conda create -n rkhs python=3.8
conda activate rkhs
pip install -r requirements.txt
params:
Since we sample y coordinate using a discrete distribution, ydes_1 and ydes_2 are the two choices for the desired y-coordinates of the obstacle.
ydes_1 - desired y-coordinate
ydes_2 - desired y-coordinate
prob - probability associated with ydes_1
cd IROS_repo_reduced
python main_reduced.py --ydes_1 ${float} --ydes_2 ${float} --prob {float between 0 and 1}
The output of the above python code is a .mp4 file stored in the folder IROS_repo_reduced/videos/.The video shows an animation of the reduced set sample selection and the associated cross-entropy cost.
The following works on synthetically generated data but in general ,it is possible to use any dataset of your choice after certain cosmetic changes in the relevant parts of the code.
Link to sample data corresponding to step 3: https://drive.google.com/drive/folders/1zKjVDz-YaB2eF5DejDLMigOodpF_T-2Z?usp=sharing
cd IROS_repo_comparison
- To reproduce the comparison shown in the video we use the data stored in
sample_datafolder.
python visualize_sample_data_ours.py
python visualize_sample_data_baseline.py
The outputs of the above python codes are two .mp4 files stored in IROS_repo_comparison/videos/.The videos show animations of the planner output trajectories corresponding to our approach and that of [6].
- It is possible to generate fresh synthetic data by running the script
data_generation.py
python data_generation.py
The script has certain modifiable parameters to generate data.Please see the comments in the script for further information. The output data is stored in the folder data_synthetic.Each data file contains an obstacle configuration and collection of possible trajectories of the obstacle.
Once the data is generated , run:
python comparison.py
This will run the optimizers for [6] and our approach on all the data files in data_synthetic folder and the resulting output files,which we again call as data files,are stored in data_comparison folder. Each data file contains trajectory information of the ego vehicle corresponding to approach of [6] and our approach.
Now that we have the optimizer/planner output we can run comparisons between the [6] and our approach. We calculate the average number of collision-free trajectories for the two approaches for each data file in data_comparison folder.
python stats_comparison.py
The output of the above script is stats_overall.txt which is stored in data_comparison folder. The file contains information about the performance of the two approaches in terms of average number of collision-free trajectories.
Finally,to visualize the comparison in terms of box and violin plots run:
python plotting_box_plots.py