mcx / cheetah-trainer

Cheetah-Trainer

This is am implementation of the Automated Residual Reinforcement Learning algorithm.

Requirement

- python3.7
- tensorflow
- tf2rl
- pybullet
- gym
- wandb

How to use

• Basic Usage

python main.py --gait sine --policy TD3 --optimiser TBPSA

The gait argument represents the gait pattern; it can be line/sine/rose/triangle. The policy argument determines the RL agent; it can be SAC/TD3 The optimiser argument chooses the parameter optimisers; it can be BO/CMA/TBPSA

• More Arguments state-mode: This could change the state representation of the RL module. leg-action-mode: This could change the action representation of the RL module. optimisation-mask: This could change the parameter search space of the black-box optimiser. num-history-observation: whether used stacked states as RL observation.

Acknowledgment

** Parts of this implementation are based on tf2rl. **