Jingyu Chen
The University of Sheffield
This is the official repository for the paper: A Deep Multi-Agent Reinforcement Learning Framework for Autonomous Aerial Navigation to Grasping Points on Loads
The project is based on the simulator of Coppliasim. Thus, make sure you have installed the Coppliasim.
We use the ORCA for demonstration learning. The C++ code for ORCA can be found in MARL_transport/src. The Cython can transform the C++ code
into the python code. To complie it, run
python setup_rvo.py build_ext –-inplace
As we used the parallel training for demo-MADDPG, we need to install
python -m pip install mpi4py
To collect the demonstration data, we run
./collect_RL_drones_demo.sh
The data will be save like orca_demonstration_ep100_3agents_env10.npz
In the paper, we proposed two algorithms, learning from the demonstrated trajectories and behaviour cloning. We compare them with the MADDPG, MAPPO and the behaviour swarm optimisation (BSO).
To run the plain MADDPG,
./train_RL_drones_V0.sh
To run the attention-based MADDPG,
./train_RL_drones_V0_attention.sh
To run the demonstrated trajectories,
./train_RL_drones_demo.sh
To run the behaviour cloning,
./train_RL_drones_orca.sh
To run the MAPPO,
./train_RL_drones_MAPPO.sh
To run the curriculum learning,
./train_curriculum_wt_ddpg.sh
To run the whole transport from navigation to manipulation,
./train_RL_drones_whole_transport_hybrid.sh