Complex interaction between mutliple neural network controlled agents
- Our paper has been accepted by IEEE Control Systems Letters (L-CSS) and IEEE Conference on Decision and Control (CDC), looking forward to meeting colleagues at Milan this December!
This is the official implementation of the algorithm ReBAR proposed in the paper "Collision Avoidance Verification of Multi-agent Systems with Learned Policies". The algorithm computes relative backprojection sets to verify the collision avoidance safety of the given multiagent system.
This codebase has been tested compatible with the following environment configuration, although other configurations may also work:
- Ubuntu
22.04 - python
3.8 - torch
2.1.2 - cuda
11.8 - gurobipy
10.0.0 - pypoman
1.0.0
To install gurobipy, please refer to this page, and to obtain a free academic license, please follow instructions here.
We provide training scripts for networks we verified in the paper, namely single/double integrator RVO policy and potential field policy.
In this implementation we support verifying 2 agent RVO policies and n-agent potential field policies.
To train RVO policies, run the following command. Note because the RVO implementation we found online is non-batchifiable, first training might take some time to generate training data by running simulations in RVO.
// Single integrator
python -m model.train.train_RVO --goal_1_x 9.0 --goal_1_y 5.0 --goal_2_x 5.0 --goal_2_y 9.0 \
--save_path model/checkpoints/single_integrator \
--batch_size 32 --lr 1e-4 --weight_decay 1e-4 --epoch 20
// Double integrator
python -m model.train.train_RVO --goal_1_x 9.0 --goal_1_y 5.0 --goal_2_x 5.0 --goal_2_y 9.0 \
--save_path model/checkpoints/double_integrator \
--batch_size 32 --lr 1e-4 --weight_decay 1e-4 --epoch 20 \
--double_integ
To train potential field policies (refer to this paper for further information about the policy), run the following command:
python -m model.train.train_potential_field \
--hidden_dim_1 20 --hidden_dim_2 20 \
--save_path model/checkpoints/potential_field \
--total_agent_num 10 \
--batch_size 32 --lr 1e-4 --weight_decay 1e-4 --epoch 20
Some pretrained checkpoints can be found in model/checkpoints/.
To verify the trained policies, run the following command. To get the monte carlo sampled RBPUA, pass in the --RBPUA flag. To plot the results, pass in the --plot flag.
// Single Integrator
python run_single_integrator.py \
--lb_x 0.0 --lb_y 0.0 --ub_x 10.0 --ub_y 10.0 \
-r 1.0 --num_cs 20 \
--hidden_dim_1 10 --hidden_dim_2 10 \
--checkpoint_1 path/to/agent1/checkpoint \
--checkpoint_2 path/to/agent2/checkpoint \
--state_uncertainty 0.5 \
--steps 5 \
--RBPUA --plot
// Double Integrator
python run_double_integrator.py \
--lb_x 0.0 --lb_y 0.0 --ub_x 10.0 --ub_y 10.0 \
-r 1.0 --num_cs 20 \
--hidden_dim_1 10 --hidden_dim_2 10 \
--checkpoint_1 path/to/agent1/checkpoint \
--checkpoint_2 path/to/agent2/checkpoint \
--state_uncertainty 0.5 \
--RBPUA --plot
// Potential Field
python run_potential_field.py \
--lb_x -10.0 --lb_y -10.0 --ub_x 10.0 --ub_y 10.0 \
--num_agents 10 -r 1.0 --num_cs 20 \
--hidden_dim_1 20 --hidden_dim_2 20 \
--checkpoint_dir /dir/containing/all/agent/checkpoints \
--state_uncertainty 0.5 \
--RBPUA --plot
We thank the authors of third-party library pypoman for their impressive work.
If you find this work helpful, please consider citing our work:
@article{dong2024collision,
title={Collision Avoidance Verification of Multiagent Systems with Learned Policies},
author={Dong, Zihao and Omidshafiei, Shayegan and Everett, Michael},
journal={arXiv preprint arXiv:2403.03314},
year={2024}
}