Skip to content

Latest commit

 

History

History
67 lines (58 loc) · 3.19 KB

README.md

File metadata and controls

67 lines (58 loc) · 3.19 KB
Raw

Nowhere to Go: Benchmarking Multi-robot Collaboration in Target Trapping Environment

Collaboration is one of the most important factors in multi-robot systems. Considering certain real-world applications and to further promote its development, we propose a new benchmark to evaluate multi-robot collaboration in Target Trapping Environment (T2E). Furthermore, we present and evaluate multiple learning-based baselines in T2E, and provide insights into regimes of multi-robot collaboration. Here is the implementation.

1. Usage

We designed several algorithms based on MADDPG, MAAC, MAPPO, and IPPO. The ./envs subfolder contains environment implementations based on MPE. And the ./offpolicy and ./onpolicy subfolders contain the corresponding algorithms implementation.

2. Installation & Environment

  1. Create an environment using conda
conda create -n t2e python=3.6
conda activate t2e
pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html
  1. Install the python package
git clone https://github.com/Dr-Xiaogaren/T2E.git
cd T2E
pip install -e .
pip install -r requirements.txt

Note that there may be redundant packages in requirements.txt.

3.Train

3.1 MAPPO

cd onpolicy/scripts
chmod +x ./train_catching_both_mappo.sh
./train_catching_both_mappo.sh

3.2 IPPO

cd onpolicy/scripts
chmod +x ./train_catching_both_ippo.sh
./train_catching_both_ippo.sh

3.3 MADDPG

cd offpolicy/scripts
chmod +x ./train_catching_both_maddpg.sh
./train_catching_both_maddpg.sh

3.4 MAAC

cd offpolicy/scripts
chmod +x ./train_catching_both_maac.sh
./train_catching_both_maac.sh

3.5 Some Notes for Train

  • When setting the number of robots in each group, the total number of robots needs to be adjusted at the same time.
  • To set the speed of the robot, one needs to modify it directly in the mpe/scenarios/simple_catching_expert_both.py file.
  • Each group comes with a heuristic that needs to be set with the --step_mode parameter. The optional range is (expert_adversary, expert_both, expert_prey, none). expert_adversary and expert_prey mean that only the captor or target use the heuristic respectively. expert_both and none mean that both use or neither use heuristics.

4. Evaluate

In eval_catching_both.sh, set the --model_dir to the data directory and --load_model_ep to ep number in the name of pretrained models. If visualization is not required, remember to remove -- save_gifs. Then just run:

cd onpolicy/scripts
chmod +x eval_catching_both.sh
eval_catching_both.sh

The visualization function has been modified to scripts/render_catching_both.sh

5. Acknowledgment

Our code frameworks are mainly implemented based on MPE, MAPPO, MAAC, and MADDPG. We thank the respective authors for open-sourcing their code.