chris-arvin/cadrl

Implementation of paper "Decentralized Non-communicating Multiagent Collision Avoidance with Deep Reinforcement Learning". NO LONGER MAINTAINED. CHECK OUT CrowdNav.

CADRL

Implementation of paper Decentralized Non-communicating Multiagent Collision Avoidance with Deep Reinforcement Learning by Yu Fan Chen, Miao Liu, Michael Everett and Jonathan P. How

This library is no longer maintained. CADRL and GA3C-CADRL is also implemented in our newest library CrowdNav, which is easier to use and can be better used for benchmarking RL navigation algorithms.

Usage

Training with specified configuration on CPU:

python train.py --config=configs/model.config

Training with specified configuration on GPU:

python train.py --config=configs/model.config --gpu

Visualize the trained agent:

python visualize.py --output_dir={OUTPUT_DIR}

Implementation details

• All the training data for imitation learning/initialization of the model is generated with RVO2. Both evaluation and test are performed on the crossing scenarios.

• The kinematics can be toggled in the env.config, which gives a hard rotation constraint.