Repository associated with our RAL submission.

1. Clone this repository:

git clone https://github.com/Basant1861/MPC-MMD.git
cd MPC-MMD

2. Create a conda environment and install the dependencies:

conda create -n mpc_mmd python=3.8
conda activate mpc_mmd
pip install -r requirements.txt

3. For CARLA(version 13), follow the instructions given in the official documentation https://carla.readthedocs.io/en/0.9.13/

Step 0 Create two folders data and stats (in the synthetic_static_obs as well as in the synthetic_dynamic_obs directories) with the following directory structure:

1. data
   • gaussian_noise
     • noise_<int(noise_level*100)>
       • ts_<num_prime>
   • beta_noise
     • noise_<int(noise_level*100)>
       • ts_<num_prime>
2. stats
   • gaussian_noise
     • noise_<int(noise_level*100)>
       • ts_<num_prime>
   • beta_noise
     • noise_<int(noise_level*100)>
       • ts_<num_prime>

Here noise_level corresponds to the control noise you want to add, typically between 0.1 and 0.5, num_prime is the rollout horizon and can be any positive integer<=100, typically 20,30,40,50,60.

Once inside the synthetic_static_obs directory follow the below steps:

Step 0 Change the following line to include the path to the optimizer module in main_mpc.py, validation.py, plot_traj_video.py:

sys.path.insert(1, 'path/to/optimizer')

Step 1 To configure the obstacle scenarios you need to edit the function compute_obs_data in main_mpc.py.

Step 2 Run the following command from the root directory:

python3 main_mpc.py --noise_levels <list_of_noise_levels(float)> --num_reduced_sets <list_of_reduced_set_sizes>
 --num_obs <int> --num_prime <int> --noises <beta,gaussian or both> --acc_const_noise <float> --steer_const_noise <float> --costs <list_of_costs> 

where costs can be one or all of <mmd_opt, cvar> and noises can be one or all of <gaussian, beta>. The above command will run for default 200 obstacle configurations. This number can be changed by modifying num_configs variable in main_mpc.py

Step 3 Once Step 2 is complete, there will be data files in the corresponding folders in data-> <gaussian_noise/beta_noise> -> noise_level/ ts_<num_prime>. Now we need to calculate the statistics for the collected data. Run the following command from the root directory:

python3 validation.py --noise_levels <from Step 2> --num_reduced_sets <from Step 2> --num_obs <from Step 2> --num_prime <from Step 2>
--noises <from Step 2> --acc_const_noise <from Step 2> --steer_const_noise <from Step 2>

This will store the statistics in the corresponding locations in the stats folder

Step 4

(A) Next we plot the box plots as well as visualize the trajectories. To plot the box plots run the following command:

python3 plot_box_plots.py --noise_levels <from Step 2> --num_reduced_sets <from Step 2>
--num_obs <from Step 2> --num_prime <from Step 2> --noises <from Step 2>

Note:

1. The above command takes in a single integer value for num_obs. To plot the box plots for different number of obstacles you need to run the above command separately for each value of num_obs.
2. Currently, plot_box_plots.py generates box plots for both the costs, i.e, mmd_opt, cvar. In order to plot the box plots for a subset of these costs you need to uncomment/edit the lines corresponding to the costs that you do not want to plot the box plots for.

(B) To generate videos for the trajectories run the following command:

python3 plot_traj_video.py --num_obs <int> --num_reduced_sets <from Step 2>
--noise_levels <from Step 2> --num_prime <from Step 2> --noises <from Step 2>
--acc_const_noise <from Step 2> --steer_const_noise <from Step 2>

Once inside the synthetic_dynamic_obs directory follow the below steps:

Step 0 Change the following line to include the path to the optimizer module in main_mpc.py, validation.py, plot_traj_video.py:

sys.path.insert(1, 'path/to/optimizer')

Step 1 Run the following command from the root directory:

python3 main_mpc.py --noise_levels <list_of_noise_levels(float)> --num_reduced_sets <list_of_reduced_set_sizes>
 --num_obs <int> --num_prime <int> --noises <beta,gaussian or both> --acc_const_noise <float> --steer_const_noise <float> --costs <list_of_costs> 

where costs can be one or all of <mmd_opt, cvar> and noises can be one or all of <gaussian, beta>. The above command will run for default 200 obstacle configurations. This number can be changed by modifying num_configs variable in main_mpc.py

Step 2 Once Step 2 is complete, there will be data files in the corresponding folders in data-> <gaussian_noise/beta_noise> -> noise_level/ ts_<num_prime>. Now we need to calculate the statistics for the collected data. Run the following command from the root directory:

python3 validation.py --noise_levels <from Step 2> --num_reduced_sets <from Step 2> --num_obs <from Step 2> --num_prime <from Step 2>
--noises <from Step 2> --acc_const_noise <from Step 2> --steer_const_noise <from Step 2>

This will store the statistics in the corresponding locations in the stats folder

Step 3

(A) Next we plot the box plots as well as visualize the trajectories. To plot the box plots run the following command:

python3 plot_box_plots.py --noise_levels <from Step 2> --num_reduced_sets <from Step 2>
--num_obs <from Step 2> --num_prime <from Step 2> --noises <from Step 2>

Note:

1. The above command takes in a single integer value for num_obs. To plot the box plots for different number of obstacles you need to run the above command separately for each value of num_obs.
2. Currently, plot_box_plots.py generates box plots for both the costs, i.e, mmd_opt, cvar. In order to plot the box plots for a subset of these costs you need to uncomment/edit the lines corresponding to the costs that you do not want to plot the box plots for.

(B) To generate videos for the trajectories run the following command:

python3 plot_traj_video.py --num_obs <int> --num_reduced_sets <from Step 2>
--noise_levels <from Step 2> --num_prime <from Step 2> --noises <from Step 2>
--acc_const_noise <from Step 2> --steer_const_noise <from Step 2>

Once inside the carla folder follow the below steps:

Step 0 Change the following line to include the path to the optimizer/Python_API folder in main_carla.py:

sys.path.insert(1, 'path/to/optimizer')
sys.path.insert(1, 'path/to/PythonAPI/carla')

Deterministic baseline : Run the following command:

python3 main_carla.py --noise_levels <float> --num_reduced_set <not_relevant, put 1> --num_exps <int>
--num_obs 3 --num_prime <not_relevant, put 1> --noises <beta/gaussian> --acc_const_noise <float>
--steer_const_noise <float> --costs det --total_obs <8/10> --town <Town05/Town10HD>

CVaR baseline/Our MMD-based approach :

python3 main_carla.py --noise_levels <float> --num_reduced_set <int> --num_exps <int>
--num_obs 3 --num_prime <int> --noises <beta/gaussian> --acc_const_noise <float>
--steer_const_noise <float> --costs cvar/mmd_opt --total_obs <8/10> --town <Town05/Town10HD>