This work aims at implementing simple MPC controller for gym's Mujoco models as described in Neural Network Dynamics for Model-Based Deep Reinforcement Learning with Model-Free Fine-Tuning and build on it by adding LQR based controllers instead of using simple shooting methods. Such controllers are then applied in parallel and the stored trajectories are used to learn a general neural network policy.
This code has been tested on python3 and requires mujoco_py installed.
Please use python3 main.py
to run. Passing --load_model
will restore the previously stored policy parameters.