环境：

• Ubuntu 20.04.6 LTS, kernel 5.15.0-139
• RTX 4090D 24GB, NVIDIA driver 580.126.09

安装步骤：

=== 0. conda 通道设置(关键!) ===

中国大陆用清华镜像,海外用官方 https://conda.anaconda.org/conda-forge

conda config --add channels defaults
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge
conda config --set channel_priority strict

验证:channels 列表第一项必须是 conda-forge(或它的镜像)

=== 1. 建 env ===

conda create -n isaaclab python=3.10 -y
conda activate isaaclab

验证 libgcc/python 来自 conda-forge:

conda list 2>&1 | grep -iE "^python |libgcc " 应该看到 Channel 列是 conda-forge

=== 2. 装 PyTorch 2.5.1 cu121(driver 580 兼容,Isaac Lab 2.1 要求) ===

pip install torch2.5.1 torchvision0.20.1 --index-url https://download.pytorch.org/whl/cu121

=== 3. 装 Isaac Sim 4.5.0(用 extras 拉齐 extscache) ===

pip install “isaacsim[all,extscache]==4.5.0” --extra-index-url https://pypi.nvidia.com

=== 4. 预修两个已知坑(在 isaaclab.sh --install 前做) ===

4a) pandas 源码 build 会拉旧 numpy 编译失败 → 强制装二进制 wheel

pip install “pandas>=2.0,❤️.0” --only-binary=:all:

4b) flatdict==4.0.1 setup.py 用了老 pkg_resources;setuptools>=80 不再带 pkg_resources

pip install “setuptools<80” wheel
pip install flatdict==4.0.1 --no-build-isolation

=== 5. 装 Isaac Lab 2.1.0 ===

cd /path/to/IsaacLab # 应已 git clone --branch v2.1.0
./isaaclab.sh --install

测试命令：

./isaaclab.sh -p scripts/demos/h1_locomotion.py 第一次启动需要等待较长时间，后续再启动速度很快