A deep reinforcement learning algorithm is used to solve the reacher benchmark problem.

The Branching Dueling Q-Network (BDQ) is a branching variant of the Dueling Double Deep Q-Network (Dueling DDQN). The code was adapted (and simplified) from here. The paper can be found here.

2 DOF reacher after 20,000 training episodes:

6 DOF reacher after 100,000 training episodes:

python 3.7.3

gym 0.9.1

tensorflow 1.14.0

mujoco_py 0.5.7

git clone https://github.com/PierreExeter/BDQ_for_reacher.git

Run the script train_continuous.py

Open the script enjoy_continuous.py

Select the trained model from the trained_models folder.

Run the script enjoy_continuous.py