Skip to content

XuShenLZ/conflict_rez

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

59 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

conflict_rez

Conflict resolution for multiple vehicles in confined spaces

Xu Shen ([email protected]), Francesco Borrelli

Project Webpage: https://bit.ly/rl-cr

rl-control-side-by-side - 2022-11-01 10-44-17 00_00_00-00_00_30

Install

  1. Clone this repo
  2. Run pip install -e . in the root level of this repo (A virtualenv is recommended). However, due to some buggy dependencies between package versions, you need to manually switch the following packages to specific versions (while ignoring the warning given by the pip dependency resolver):
    1. pip install pettingzoo==1.20.1
    2. pip install supersuit==3.5.0
    3. pip install stable-baselines3==1.6.0
    4. pip install gym==0.25.0

Testing

  1. A pretrained model can be downloaded here.
  2. Run python confrez/rl/experiment.py to see the steps taken by the DQN policy to resolve the conflict in the discrete environment.
  3. Run python confrez/rl/record_states_history.py to generate a .pkl file that records the steps taken by the RL agents, which will serve as the configuration strategies for the trajectory planning problems.
  4. Run python confrez/control/vehicle.py to plan a single-vehicle collision free trajectory following the strategy-guided configurations.
  5. (Centralized method) Run python confrez/control/multi_vehicle_planner.py to solve the multi-vehicle trajectory planning problem to resolve conflict.
  6. (Distributed method) Run python confrez/control/vehicle_follower.py so that each vehicle generates its own strategy-guided reference trajectory, and then follows it with distributed MPC to avoid collisions.

Training

  1. Run python confrez/rl/train.py to train a new policy.
  2. You may set the random_reset argument to True so that you train the policy with a random subset out of 4 vehicles, which may lead to a more generalizable policy. But it also require longer training time.
  3. Different random seeds and different parameter tuning will lead to different behaviors in the trained policy. With your new policy, the vehicles may (almost for sure) take different actions and resolve the conflict in another way.