A1_quadruped_env

Code for paper:

"Identifying Important Sensory Feedback for Learning Locomotion Skills"

The DRL framework trains A1 quadruped robot to perform five locomotion skills:

• Motor skills trained with full states: balance recovery, trotting, and bounding.

• Motor skills trained with key states which are identified through the proposed approach in this paper: balance recovery, trotting, bounding, pacing, and galloping.

More details can be found in the paper.

Install dependencies

The code has been tested in Ubuntu 16.04, 18.04 and 20.04. Recommend to install Anaconda (link for install instructions here)

After installing Anaconda, set up conda environment and install dependencies as below:

./run_create_conda_env.sh

Run trained policies

Full-state policies

• Trained full-state policies can be found in the corresponding folder under: A1_quadruped_env/full_state/SAC/SAC/SAC/record/a1/

• To run the policies:

  cd A1_quadruped_env/full_state/SAC/SAC/

  python run_a1_3D_SAC_standup.py

  python run_a1_3D_SAC_trot.py

  python run_a1_3D_SAC_bound.py

• To train new policies with full states under the same folder:

  python train_a1_3D_SAC_standup.py

  python train_a1_3D_SAC_trot.py

  python train_a1_3D_SAC_bound.py

Key-state policies

• Trained key-state policies can be found in the corresponding folder under: A1_quadruped_env/key_state/SAC/SAC/SAC/record/a1/

• To run the policies:

  cd A1_quadruped_env/key_state/SAC/SAC/

  python run_a1_3D_SAC_standup.py

  python run_a1_3D_SAC_trot.py

  python run_a1_3D_SAC_bound.py

  python run_a1_3D_SAC_pace.py

  python run_a1_3D_SAC_gallop.py

• To train new policies with key states under the same folder:

  python train_a1_3D_SAC_standup.py

  python train_a1_3D_SAC_trot.py

  python train_a1_3D_SAC_bound.py

  python train_a1_3D_SAC_pace.py

  python train_a1_3D_SAC_gallop.py