final_project

基于mujoco playground的go2 毕设

环境搭建

环境建议 在以下环境中完成了测试

• 操作系统：Ubuntu 22.04/20.04 LTS x64
• NVIDIA-SMI: >= 12.8
• CUDA：12/11.8
• CUDNN: 8.9

git clone --recurse-submodules https://github.com/120090162/final_project.git
cd final_project

依赖安装

conda create -n cimpc-rl python=3.12 -c conda-forge
conda activate cimpc-rl
pip install uv

cd tools/mujoco_playground

uv pip install -U "jax[cuda12]"==0.5.3
python -c "import jax; print(jax.default_backend())"   # 打印信息为gpu

uv pip install -e ".[all]"
uv pip install -U "jax[cuda12]"==0.5.3 # 避免jax依赖问题
python -c "import mujoco_playground"  # 开始下载MuJoCo Menagerie库

训练

• 无可视化

  使用例子

  sudo apt install ffmpeg 
  # make sure your are in the root of project
  python examples/train_jax_ppo.py --env_name CartpoleBalance

• 可视化

  安装rscope可视化

  cd tools/rscope
  uv pip install -e ".[all]"

  使用例子

  # make sure your are in the root of project
  python examples/train_jax_ppo.py --env_name PandaPickCube --rscope_envs 16 --run_evals=False --deterministic_rscope=True
  # In a separate terminal
  python -m rscope
  
  # 如果卡死
  pkill -9 -f rscope

定义新环境

• 训练

./scripts/run_train.sh \
    --env_name=Go2JoystickFlatTerrain \
    --use_wandb=True \
    --log_training_metrics=True \
    --run_evals=False

• 单纯可视化

./scripts/run_train.sh \
    --env_name=Go2JoystickFlatTerrain \
    --play_only=True \
    --load_checkpoint_path=logs/<env_name>/<timestamp>/checkpoints \
    --num_videos=<video_num_to_render>

测试

# sim to sim
uv pip install pygame

python sim2sim/play_go1_keyboard.py --policy_name=<_policy前面的名称>
python sim2sim/play_go2_keyboard.py --policy_name=<_policy前面的名称>

工具

# convert brax model to onnx type
uv pip install "tensorflow-cpu>=2.19.0" "tf2onnx>=1.16.1" "onnx>=1.16.0" onnxruntime "ml-dtypes==0.5.4" "numpy==1.26.4"
python utils/brax_to_onnx.py \
    --checkpoint_path=logs/<env_name>/<timestamp>/checkpoints/<step_num> \
    --env_name=<env_name>

# convert urdf to mjcf
uv pip install urdf2mjcf

python utils/urdf2xml.py --urdf_path=sample/robot.urdf

# visulize
python examples/show_mjcf.py --mjcf_path=assets/sample/robot.xml