viveksharmaaa / NNRCCM

Learning Tube Certified Control Using Robust Contraction Metric

We modified code taken from C3M which is the repo open-sourced by the authors of "Learning Certified Control Using Contraction Metric"

Install package dependencies

Dependencies include torch, tqdm, numpy, and matplotlib. You can install them using the following command.

pip install -r requirements.txt

Usage

The script main.py can be used for learning the controller. Usage of this script is as follows

usage: main.py [-h] [--task TASK] [--bs BS]
               [--num_train NUM_TRAIN] [--num_test NUM_TEST]
               [--lr LEARNING_RATE] [--epochs EPOCHS] [--lr_step LR_STEP]
               [--lambda _LAMBDA] [--w_lb W_LB] [--log LOG]
               [--sigma _SIGMA]

optional arguments:
  -h, --help            show this help message and exit
  --task TASK           Name of the model.
  --bs BS               Batch size.
  --num_train NUM_TRAIN
                        Number of samples for training.
  --num_test NUM_TEST   Number of samples for testing.
  --lr LEARNING_RATE    Base learning rate.
  --epochs EPOCHS       Number of training epochs.
  --lr_step LR_STEP
  --lambda _LAMBDA      Convergence rate: lambda
  --w_lb W_LB           Lower bound of the eigenvalue of the dual metric.
  --log LOG             Path to a directory for storing the log.
  --sigma _SIGMA        Disturbance bound

Jointly learn the controller and RCCM

Run the following command to learn a robust control contraction metric (RCCM), controller while minimizing gain for PVTOL model.

mkdir log_PVTOL
python main.py --log log_PVTOL --task PVTOL --sigma 1
python main.py --log log_QUADROTOR_9D --task QUADROTOR_9D  --sigma 1
python main.py --log log_NL --task NL --sigma 1
python main.py --log log_TLPRA --task TLPRA --sigma 1 --lambda 2

Please note that for generating tube size for variables of your choice for the system PVTOL, you may modify the matrices C,D,C_ref,D_ref in systems\system_PVTOL.py

Comparisons

Run the following command to reproduce plot comparison results between NN-RCCM-P (ours) and NN-CCM.

python plot.py --pretrained_CCM log_PVTOL/controller_best_CCM.pth.tar --pretrained_RCCM log_PVTOL/controller_best_ref.pth.tar --task PVTOL --plot_type mean_std_plot_pos --sigma 1 --plot_dims_pos 0 1 --nTraj 100
python plot.py --pretrained_CCM log_QUADROTOR_9D/controller_best_CCM.pth.tar --pretrained_RCCM log_QUADROTOR_9D/controller_best_ref.pth.tar --task QUADROTOR_9D --plot_type mean_std_plot_pos --sigma 1 --plot_dims_pos 0 1 2 --nTraj 100
python plot.py --pretrained_CCM log_NL/controller_best_CCM.pth.tar --pretrained_RCCM log_NL/controller_best_ref.pth.tar --task NL --plot_type mean_std_plot_pos --sigma 1 --plot_dims_pos 0 1 2 --nTraj 100
python plot.py --pretrained_CCM log_TLPRA/controller_best_CCM.pth.tar --pretrained_RCCM log_TLPRA/controller_best_ref.pth.tar --task TLPRA --plot_type mean_std_plot_pos --sigma 1 --plot_dims_pos 0 1 --nTraj 100

For demonstration of tube-based safe motion planning from Section 4.2.2 in the paper, run the following file from the self-contained folder motion-planning in MATLAB 2022b

plot_compare_trajs.m

The motion planner uses trajectory-optimization toolbox taken from OptimTraj