Learning Multi-Agent Policies using GNNs

Dependencies

• Python 3
• OpenAI Gym
• PyTorch
• Gym-Flock

Available algorithms:

• Behavior Cloning as described in ArXiv python3 train.py cfg/cloning.cfg
• DAGGER imitation learning python3 train.py cfg/dagger.cfg
• Deep Deterministic Policy Gradients (TODO)

Other code:

• python3 flocking_gym_test.py provides test code for the Gym Flock environments

To Do

• Plot centralized and decentralized baselines
• Obstacle avoidance
• Follow 1 or 2 leaders
• Play with control costs
• DDPG