RL-MPC-Quadrupedal-Controller-For-Rough-Terrain-Locomotion

The MPC controller is based on the paper titled Dynamic Locomotion in the MIT Cheetah 3 Through Convex Model-Predictive Control and it's implementation is taken from here.

Running the code

1. First clone the repository:

   git clone 
   

2. Compile the MPC controller using:

   cd quadruped_ctrl/
   mkdir build
   cd build
   cmake ..
   make -j 4
   

   This creates a .so file inside the build directory named libquadruped_ctrl.so.

3. In order to train the robot or to run robot on a pretrained model, run the main.py file.

   python main.py
   

Configs

There are two different configs that determine the behavior of the simulation. The first one is config/training_params.yaml. This file contains the training parameters.

The second config file is quadruped_ctrl/config/quadruped_ctrl_config.yaml. This file contains all the simulation related configurations. The values of interest in this config are: visualization and terrain.

The visualization should be set to False when the display needs to be disabled. This is useful during training.

The terrain option specifies the terrain the robot will walk on. The two options are plane and hybrid.

References

1. https://github.com/mit-biomimetics/Cheetah-Software
2. https://dspace.mit.edu/bitstream/handle/1721.1/138000/convex_mpc_2fix.pdf
3. https://github.com/Derek-TH-Wang/quadruped_ctrl
4. https://github.com/Derek-TH-Wang/hybrid_gait