MiniCheetahEnv

• a pybullet-gym environment for Mini Cheetah

To Do:

• Check MPC implementation of pybullet, and the simulation bed configuration.
• Import Mini Cheetah in the place of Laikago, do the requied system indentification and test the MPC controller.
• Clean and develop the simulation bed into a gym env with approproate functions and classes.
• Build a independent Domain Randomizer class, to work hand in hand with the env.
• Integrate, test and verify env.
• Add functions for capturing image as the observation.
• Make it a gym package.

Suggestions:

• Addition of docstrings.
• Implement utils and logger files for performace tracking and comparison .
• Add DR for the images aswell (if required).
• Add multi threading / make vectorized env for paralelized training.
• Add stable baselines support for training.

MPC controller

Installtion:

• Install the motion_imitation repository and all the requirements as per the instructions given here.
• Replace the existing "mpc_controller" folder with the folder in this repository.

Run:

• To run the mini cheetah mpc controller(untuned) demo,

  cd mpc_controller
  python locomotion_controller_example.py
  

References

Papers:

1. From Pixels to Legs: Hierarchical Learning of Quadruped Locomotion. Paper, CoRL Presentation
2. Vision-aided Dynamic Exploration of Unstructured Terrain(in Mini Cheetah). Paper, Video

Paper and Code:

1. Previous Implementations - Github
2. MPC Controller in Pybullet
3. SlopedTerrainLinearPolicy(for DR)
4. Sim-to-Real: Learning Agile Locomotion For Quadruped Robots(DR implementations of minatour)
5. Active Domain Randomization