Providing Safety Assurances for Systems with Unknown Dynamics

arxiv link: https://arxiv.org/abs/2403.05771

TurtleBot experiment supplemental video link: https://youtu.be/D6ZxZG0rQ_U

Important Note: Please remember to add all the files in this repository to your MATLAB path.

Inverted Pendulum Experiment

• The experiment directory contains
  • Datasets for training ensemble dynamics models in dyn_datasets
  • A trained ensemble dynamics model using 100 training samples.
  • Ensemble forward results for the ensemble model trained with 100 training samples. Located in directory ensemble_dyn_forward_results
  • A grid for helperOC used in BRT computation inv_pend_pi_10_200_200_grid_data.mat
  • BRT figures for the ground truth, 0std (Baseline 1 Mean Dynamics), 3std (our method) located in directory figs
• To visualize BRT figures for this experiment (Fig. 1 in the paper)
  • Run brt_plotting.m in MATLAB
• To generate your datasets
  • Run python inv_pend_dyn_data_collection.py
    • You can change the parameters of the pendulum l,m,b and the number of training samples within the script. The dataset will be saved to dyn_datasets.
• To traing your own ensemble dynamics model and generate the ensemble forward results
  • Run python inv_pend_dyn_training_driver.py
    • You can change training arguments, such as number of ensemble models, number of hidden layers, and learning rate, in inv_pend_dyn_training_config.txt.
  • Run python inv_pend_ensemble_dyn_forward.py
    • You need to make sure the training experiment name is properly spelled out within the script
• To compute the BRT (after generating the ensemble forward results)
  • Run inv_pend_brt_comp_script.m in MATLAB. Make sure the dataset name and forward result file name are correct within the script.
  • You can save the figure and use brt_plotting.m to compare BRTs of different dynamics models, model uncertainty level, etc.

Last updated 3/14/24