RANS-RRT*

This is an implementation of our work "Risk-Averse RRT* Planning with Nonlinear Steering and Tracking Controllers for Nonlinear Robotic Systems Under Uncertainty".

RANS-RRT* (Risk-Averse Nonlinear Steering RRT*) is a sampling-based planning algorithm based on RRT*. The current implementation uses unicycle dynamics for the nonlinear steering function. The generated trajectory can be passed down to low-level tracking controllers: LQR, LQR with multiplicative noise terms, and nonlinear model predictive control (NMPC).

Authors:

• Sleiman Safaoui - UT Dallas
• Ben Gravell - UT Dallas
• Venkatraman Renganathan - UT Dallas
• Dr. Tyler Summers - UT Dallas

Table of Contents

• Cloning the Package
• Dependencies
• Package Organization and Description
• Using the Package
• License

Cloning the Package

The package contains submodules. Please clone with:

git clone https://github.com/TSummersLab/RANS-RRTStar.git --recurse

Note: When using GitHub desktop, the application initializes all submodules when cloning for the first time. If a submodule is not initialized after pulling changes, please use the Git Bash tool or terminal and run git submodule update --init at the root of the repository.

Dependencies

The package uses Python 3 and requires the following packages:

• NumPy
• SciPy
• CasADi
• namedlist

Package Organization and Description

The main scripts are organized as follows:

config.py
rans_rrtstar.py
get_opt_paths.py
scripts
    |__monte_carlo.py
saved_data
monte_carlo

Config File

config.py provides a list of stetting that can be changed by the user (all are documented in the file). We note the following parameters:

• NUMSAMPLES: Number of random samples that RRT* will try to add. Not all will be added to the tree.
• STEER_TIME: Number of discrete time steps used for steering between two nodes.
• RRT: Switches between RRT (not asymptotically optimal) and RRT* (asymptotically optimal) implementations
• DRRRT: Switches between using a distributionally robust collision check and not using it.

RRT	DRRT	Resulting Algorithm
True	False	RRT (with nonlinear dynamics)
True	True	DR-RRT (with nonlinear dynamics)
False	False	RRT* (with nonlinear dynamics)
False	True	RANS-RRT*

• ENVNUM: Index for the environment. We provide five environments. More can be added following the same format. -SAVEPATH: Directory for saving data relative to the root of the package. We recommend storing the results in new directories in saved_data.
• SIGMAW: Process noise covariance. Must be balanced with BETA (see below).
• SIGMAV: Measurement noise covariance. We do NOT use this. Keep all zeros.
• CROSSCOR: Cross correlation between process and measurement noise covariance. We do NOT use this. Keep all zeros.
• BETA: Risk bound (% of experiments that can fail). This is an upper bound. The smaller it is, the larger the tightening.
• TMAX: Maximum number of time steps used for planning.

RANS-RRT* File

rans_rrtstar.py contains all the code for RANS-RRT* (data structures, functions and methods for the planner, and functions for storing the resulting data). After this script runs, two files are stored in the directory specified in the config file: * the node list data (RANS-RRT* tree): NodeListData_v<version>_<timestamp> (pickle file) * an image of the generated tree: plot_tree_v<version>_<timestamp> (png)

Post-Processing Trajectory File

get_opt_paths.py takes tree generated by rans_rrtstar.py and extracts the optimal trajectory from the root to the goal. It saves the inputs and states corresponding to that trajectory, and a png image of it, in the save directory:

• OptTraj_v*_inputs (optimal trajectory inputs)
• OptTraj_v*_states (optimal trajectory states)
• OptTraj_v*_plot.png (optimal trajectory image)

Furthermore, it shortens the trajectory by replanning between successive RANS-RRT* nodes using a dynamic number of discrete time steps (instead of using STEER_TIME steps). The shortened trajectory is also checked to satisfy the distributionally robust collision check. The shortened trajectory inputs, states, and image are also saved in the same directory as:

• OptTraj_short_v*_inputs (shortened optimal trajectory inputs)
• OptTraj_short_v*_states (shortened optimal trajectory states)
• OptTraj_short_v*_plot.png (shortened optimal trajectory image)

Monte Carlo File

scripts/monte_carlo takes the post-processed data from get_opt_paths.py and performs tracking simulations with realized noise. The Monte Carlo results are stored in the directory specified by the MC_FOLDER variable. Each experiment setup (new environment, new noise distribution, ...), has a unique UNIQUE_EXP_NUM corresponding to all the data in the inputs file name excluding the words "OptTraj_" and "_inputs". For that experiment setup, each experiment run (with its own noise realizations) will be stored in a folder named by a 12-digit number (e.g.000000000001). These folders are automatically generated based on the num_trials and trials_offset (see later).

In main of the script, the following variables can be modified.

• The trajectory specified by the chosen input file input_file (e.g. OptTraj_short_v2_0_1623778273_inputs).
• The noise distribution is specified using the noise_dist variable. The available distributions are:
  • nrm: Gaussian/normal distribution
  • lap: Laplacian distribution
  • gum: Gumbel distribution
• The desired low-level controllers specified in the controller_str_list list. The available controllers are:
  • 'open-loop': open loop control
  • 'lqr': LQR
  • 'lqrm': LQR with multiplicative noise terms
  • 'nmpc': nonlinear MPC
• The number of trials or experiment runs to perform: num_trials.
• To allow for running batches of experiments then collecting all of their data, the number of completed experiments can be specified in trials_offset.
  • num_trials = 2, trials_offset = 0 generates runs stored in folders 000000000001, 000000000002.
  • num_trials = 3, trials_offset = 1 generates runs stored in folders 000000000002, 000000000003, 000000000004.
  • Existing folders are overwritten.
• To generate new experiments set run_flag to True.
  • To collect the data of already existing experiments:
    • Set num_trials to the available experiments
    • Set trials_offset to the number of the first experiment
    • Set run_flag to False Note: all integers between trials_offset and trials_offset + num_trials must exist. Otherwise, an error is thrown.

Saved Data Directory

saved_data: Preferable directory to store data.

Monte Carlo Results Directory

monte_carlo: Directory where Monte Carlo results are stored.

Using the Package

To use the package follow these steps.

1. Update the config file (config.py).
2. Run the RANS-RRT* file (rans_rrtstar.py).
   • this generates the trajectory and saves two files
3. Copy the name of the node list (e.g. NodeListData_v2_0_1623778273).
4. Update the filename variable in main of get_opt_paths.py with the copied node list name.
   • this extract the optimal trajectory and shortens it (in time) and saves six files
5. Copy the name of the trajectory's inputs (e.g. OptTraj_short_v2_0_1623778273_inputs or OptTraj_v2_0_1623778273_inputs).
6. Update the input_file variable in main of monte_carlo.py with the copied input file name.
7. Update the remaining configuration variables in monte_carlo.py
   • input_file at the top of the script
   • noise_dist, controller_str_list, num_trials, trials_offset, and run_flag in main
8. Run the Monte-Carlo script monte_carlo.py,
   • this will generate the experiment runs in monte_carlo/env<UNIQUE_EXP_NUM>/<trials_offset+1> through monte_carlo/env<UNIQUE_EXP_NUM>/<trials_offset+num_trials>.
   • it will also save the trajectory images in monte_carlo/path_plots/env<UNIQUE_EXP_NUM>.
   • statistics about the results are printed.

License

See LICENSE file.