Path-following and Collision Avoidance Environment for DRL Control

This repo implements a 6-DOF simulation model for an AUV according to the stable baselines (OpenAI) interface for reinforcement learning control. The environment contains a 3D path, obstacles and an ocean current disturbance. The goal for the agents is to steer the AUV on-path while combating disturbances and avoid obstacles along the trajectory.

Getting Started

To install all packages needed in your virtual environment, run:

conda env create -f environment.yml

Training an agent:

All hyperparameters and setup can be tuned in the file train.py and init.py.

For training an agent, run:

python train.py --exp_id [x]

Where x is the experiment id number.

Running an agent in the environment

For running an agent in any scenario, use:

python run.py --exp_id [x] --scenario [scenario] --controller_scenario [controller_scenario] --controller [y]

Where x is the experiment id number, scenario is what scenario to run, controller_scenario is which scenario the controller was trained in and y is which agent number to run. If no y is provided, the agent called "last_model.pkl" is chosen. Scenarios can be either of "beginner", "intermediate", "proficient", "advanced", "expert", "test_path", "test_path_current" (Path following with disturbance), "horizontal", "vertical" or "deadend".