APRL

Code to replicate Efficient Real-World RL for Legged Locomotion via Adaptive Policy Regularization. This repo contains code for training a simulated or real Go1 quadrupedal robot to walk from scratch. This code has been tested on Ubuntu 18.04 LTS with Python 3.10.

Installation

The tested Python version is 3.10, create a new environment with conda create -n aprl python=3.10 and activate it with conda activate aprl.

First copy the lib directory of Unitree SDK to the unitree_go1_wrapper/lib.

Then clone down this DMCGym repo and install it with pip install -e .. You may need to remove the first line of requirements.txt in that repo and change it as follows:

# Before the change
gym[mujoco] >= 0.21.0, < 0.24.1

# After the change
gym >= 0.21.0, < 0.24.1

Then run the following in this repo to install the dependencies:

pip install -r requirements.txt
pip install -e .

Special note with Windows users: Jax does not offer official cuda support for Windows binary wheels. However, you can use instructions in this experimental repo to install a cuda-enabled version of Jax. Our code uses jax[cuda111]==0.3.25. Make sure that the jax and jaxlib packages are both installed with the correct version. Also make sure that your nvidia driver shows support for cuda 11 when you run nvidia-smi.

Setting up the real robot

Real Robot Setup

Training

Example command to run real training

cd training/

MUJOCO_GL=egl CUDA_VISIBLE_DEVICES=0 MUJOCO_EGL_DEVICE_ID=0 \
XLA_PYTHON_CLIENT_PREALLOCATE=false python train_online.py \
  --env_name=Go1SanityReal-Empty-SepRew-v0 \
  --save_buffer=True \
  --load_buffer \
  --utd_ratio=20 \
  --start_training=1000 \
  --config=configs/droq_config.py \
  --config.critic_layer_norm=True \
  --config.exterior_linear_c=12.0 \
  --config.target_entropy=-12 \
  --save_eval_videos=False \
  --eval_interval=-1 \
  --save_training_videos=False \
  --training_video_interval=5000 \
  --eval_episodes=1 \
  --max_steps=40000 \
  --log_interval=1000 \
  --save_interval=10000 \
  --seed=0 \
  --project_name=APRL_real_reproduce \
  --tqdm=True \
  --save_dir=saved_real_exp \
  --task_config.action_interpolation=True \
  --task_config.enable_reset_policy=True \
  --task_config.Kp=20 \
  --task_config.Kd=1.0 \
  --task_config.limit_episode_length=0 \
  --task_config.action_range=1.0 \
  --task_config.frame_stack=0 \
  --task_config.action_history=1 \
  --task_config.rew_target_velocity=1.5 \
  --task_config.rew_energy_penalty_weight=0.0 \
  --task_config.rew_qpos_penalty_weight=2.0 \
  --task_config.rew_smooth_torque_penalty_weight=0.005 \
  --task_config.rew_pitch_rate_penalty_factor=0.4 \
  --task_config.rew_roll_rate_penalty_factor=0.2 \
  --task_config.rew_joint_diagonal_penalty_weight=0.00 \
  --task_config.rew_joint_shoulder_penalty_weight=0.00 \
  --task_config.rew_joint_acc_penalty_weight=0.0 \
  --task_config.rew_joint_vel_penalty_weight=0.0 \
  --task_config.center_init_action=True \
  --task_config.rew_contact_reward_weight=0.0 \
  --action_curriculum_steps=30000 \
  --action_curriculum_start=0.35 \
  --action_curriculum_end=0.6 \
  --action_curriculum_linear=True \
  --action_curriculum_exploration_eps=0.15 \
  --task_config.filter_actions=8 \
  --reset_curriculum=True \
  --reset_criterion=dynamics_error \
  --task_config.rew_smooth_change_in_tdy_steps=1 \
  --threshold=1.5

Example command to run simulated training

cd training/

MUJOCO_GL=egl CUDA_VISIBLE_DEVICES=0 MUJOCO_EGL_DEVICE_ID=0 \
XLA_PYTHON_CLIENT_PREALLOCATE=false python train_online.py \
  --env_name=Go1SanityMujoco-Empty-SepRew-v0 \
  --save_buffer=True \
  --load_buffer \
  --utd_ratio=20 \
  --start_training=1000 \
  --config=configs/droq_config.py \
  --config.critic_layer_norm=True \
  --config.exterior_linear_c=12.0 \
  --config.target_entropy=-12 \
  --save_eval_videos=False \
  --eval_interval=-1 \
  --save_training_videos=False \
  --training_video_interval=5000 \
  --eval_episodes=1 \
  --max_steps=40000 \
  --log_interval=1000 \
  --save_interval=10000 \
  --seed=0 \
  --project_name=APRL_sim_reproduce \
  --tqdm=True \
  --save_dir=saved_sim_exp \
  --task_config.action_interpolation=True \
  --task_config.enable_reset_policy=False \
  --task_config.Kp=20 \
  --task_config.Kd=1.0 \
  --task_config.limit_episode_length=0 \
  --task_config.action_range=1.0 \
  --task_config.frame_stack=0 \
  --task_config.action_history=1 \
  --task_config.rew_target_velocity=1.5 \
  --task_config.rew_energy_penalty_weight=0.0 \
  --task_config.rew_qpos_penalty_weight=2.0 \
  --task_config.rew_smooth_torque_penalty_weight=0.005 \
  --task_config.rew_pitch_rate_penalty_factor=0.4 \
  --task_config.rew_roll_rate_penalty_factor=0.2 \
  --task_config.rew_joint_diagonal_penalty_weight=0.00 \
  --task_config.rew_joint_shoulder_penalty_weight=0.00 \
  --task_config.rew_joint_acc_penalty_weight=0.0 \
  --task_config.rew_joint_vel_penalty_weight=0.0 \
  --task_config.center_init_action=True \
  --task_config.rew_contact_reward_weight=0.0 \
  --action_curriculum_steps=30000 \
  --action_curriculum_start=0.35 \
  --action_curriculum_end=0.6 \
  --action_curriculum_linear=True \
  --action_curriculum_exploration_eps=0.15 \
  --task_config.filter_actions=8 \
  --reset_curriculum=True \
  --reset_criterion=dynamics_error \
  --task_config.rew_smooth_change_in_tdy_steps=1 \
  --threshold=1.5