Quadruped Locomotion Reinforcement Learning

This is an environment for learning quadrupedal locomotion depending on either bullet physics engine or mujoco and dm_control (not implemented yet), containing

• Interfaces of quadruped states, creating terrain from elevation maps and environmental states
• Motor identification, i.e. Actuator Network
• Curriculum learning on disturbance and terrain
• Abundant command, force, torque, terrain and trajectory visualization

tianshou is used for training in our examples.

Requirements: python>=3.7, pybullet>=3.2, torch>=1.10, numpy, scipy, wandb

For train, run:

PYTHONPATH=./ python example/loct/train.py --task example/lctv0.yaml \
  --batch-size 8192 --repeat-per-collect 8 --step-per-epoch 160000 \
  --norm-adv 1 --run-name example