LGPF: Language-Guided Pattern Formation

Installation & Setup

Platform and Dependencies:

• Ubuntu 20.04
• Python 3.7
• Pytorch 1.12.1
• OpenAI gym 0.26.2 (https://github.com/openai/gym)

1. Install the improved MPE:

cd multiagent-particle-envs
pip install -e .

2. Install the specified version of Pytorch

• Ref: Installation Instruction

3. Install other required packages

pip install -r requirements.txt

Model Training

cd maddpg
python main_vec.py --exp_name simspread_n10 --scenario simple_spread_nvar --critic_type gcn_max --cuda --eval_freq 1000 --num_steps 75 --num_episodes 60000 --num_agents 10 --num_seeds 3 --tau 0.005 --force

Model Evaluation

Simulation

cd maddpg
# Evaluate the custom-trained ckpt 
python inference.py --exp_name simspread_n10 --scenario simple_spread_nvar --critic_type gcn_max --cuda --num_steps 75 --num_agents 10 --num_seeds -1

# Evaluate the provided ckpt
python inference.py --save_dir <ckpt_dir> --exp_name coop_navigation_n20 --scenario simple_spread_custom_n20 --critic_type gcn_max --cuda --num_steps 100

Real robots

cd maddpg
export PYTHONPATH=<path to maddpg directory>
python real_deploy/inference_real.py --save_dir <ckpt_dir> --exp_name <exp_name> --scenario test --critic_type gcn_max --cuda --num_steps 100 --num_agents 10 --shape <shape>

• <ckpt_dir> should be exp_ckpt if you are using the checkpoints provided by this repo
• <exp_name> can be coop_navigation_dire_n10, coop_navigation_vel0_n10
• The available options for <shape> can be found in maddpg/real_deploy/tools.py

The differences between these setups are: (see TODO 2-2 in eval_real.py)

• coop_navigation_dire_n10 use the get_direction(curr_agent.velocity) to compute the first two element of the observation
• coop_navigation_vel0_n10 use the np.array([0.0, 0.0]) as the first two element of the observation

Acknowledgement

The MADDPG code is based on the DDPG implementation of https://github.com/ikostrikov/pytorch-ddpg-naf

The improved MPE code is based on the MPE implementation of https://github.com/openai/multiagent-particle-envs

The GCN code is based on the implementation of https://github.com/tkipf/gcn

The PIC code is based on the implementation of https://github.com/IouJenLiu/PIC

License

LGPF is licensed under the MIT License